fix: Bedrock: Mantle models template api on ${AWS_REGION}, but aren't served in every region

Co-authored-by: opencode-agent[bot] <opencode-agent[bot]@users.noreply.github.com>

.github/workflows/issue-fixer.yml

@@ -18,9 +18,7 @@ jobs:
     if: >-
       github.repository == 'anomalyco/models.dev'
       && !contains(github.event.issue.labels.*.name, 'provider:openai')
-      && !contains(github.event.issue.labels.*.name, 'provider:pioneer')
       && github.event.client_payload.provider != 'openai'
-      && github.event.client_payload.provider != 'pioneer'
     runs-on: ubuntu-latest
     env:
       GH_TOKEN: ${{ github.token }}

.github/workflows/pr-reviewer.yml

@@ -7,6 +7,7 @@ on:
 
 permissions:
   contents: read
+  issues: write
   pull-requests: write
 
 concurrency:
@@ -22,6 +23,19 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
+      - name: Clear ready label
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          READY_LABEL: "reviewer: ready"
+        run: |
+          set -euo pipefail
+          gh label create "$READY_LABEL" --repo "$GITHUB_REPOSITORY" --color "0E8A16" --description "Automated review found no actionable items" --force
+          labels="$(gh pr view "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --json labels --jq '.labels[].name')"
+          if grep -Fxq "$READY_LABEL" <<< "$labels"; then
+            gh pr edit "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --remove-label "$READY_LABEL"
+          fi
+
       - name: Checkout trusted base revision
         uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5
         with:
@@ -53,15 +67,19 @@ jobs:
       - name: Run pull request reviewer
         env:
           OPENCODE_API_KEY: ${{ secrets.OPENCODE_API_KEY }}
-          OPENCODE_PERMISSION: '{"*":"deny","read":"allow","glob":"allow","grep":"allow","external_directory":"deny"}'
+          OPENCODE_PERMISSION: '{"*":"deny","read":"allow","glob":"allow","grep":"allow","mark-pr-ready":"allow","external_directory":"deny"}'
         run: |
           set -euo pipefail
           EVENTS_FILE="$RUNNER_TEMP/pr-reviewer-events.jsonl"
           RESPONSE_FILE="$RUNNER_TEMP/pr-reviewer-response.md"
+          PR_REVIEW_READY_FILE="$RUNNER_TEMP/pr-reviewer-ready"
           echo "RESPONSE_FILE=$RESPONSE_FILE" >> "$GITHUB_ENV"
+          echo "PR_REVIEW_READY_FILE=$PR_REVIEW_READY_FILE" >> "$GITHUB_ENV"
+          export PR_REVIEW_READY_FILE
+          rm -f "$PR_REVIEW_READY_FILE"
 
           opencode run --agent pr-reviewer -m opencode/grok-4.5 --format json <<'EOF' | tee "$EVENTS_FILE"
-          Review this pull request using the trusted reviewer instructions. Start with `.pr-review/pull-request.json`, `.pr-review/diff.patch`, `AGENTS.md`, and the contributing guidance in `README.md`. Read `sync.md`, the reasoning-options audit guide, schema code, and nearby base-revision files when relevant to the changed files. Use only the read, glob, and grep tools. Return only the final review comment in the agent's required output format. Never include progress narration or passed-check summaries.
+          Review this pull request using the trusted reviewer instructions. Start with `.pr-review/pull-request.json`, `.pr-review/diff.patch`, `AGENTS.md`, and the contributing guidance in `README.md`. Read `sync.md`, the reasoning-options audit guide, schema code, and nearby base-revision files when relevant to the changed files. Use only the read, glob, grep, and mark-pr-ready tools. Return only the final review comment in the agent's required output format. Never include progress narration or passed-check summaries.
           EOF
 
           if ! jq -ers 'map(select(.type == "text") | .part.text) | last | select(length > 0)' "$EVENTS_FILE" > "$RESPONSE_FILE"; then
@@ -73,4 +91,9 @@ jobs:
         env:
           GH_TOKEN: ${{ github.token }}
           PR_NUMBER: ${{ github.event.pull_request.number }}
-        run: gh pr comment "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --body-file "$RESPONSE_FILE"
+          READY_LABEL: "reviewer: ready"
+        run: |
+          gh pr comment "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --body-file "$RESPONSE_FILE"
+          if [[ -f "$PR_REVIEW_READY_FILE" ]]; then
+            gh pr edit "$PR_NUMBER" --repo "$GITHUB_REPOSITORY" --add-label "$READY_LABEL"
+          fi

.github/workflows/sync-models.yml

@@ -95,6 +95,7 @@ jobs:
         run: bun validate
 
       - name: Report changes
+        id: report
         env:
           GH_TOKEN: ${{ steps.committer.outputs.token }}
           BRANCH: automation/sync-models-${{ matrix.provider }}
@@ -119,9 +120,20 @@ jobs:
           git checkout -B "$BRANCH"
           git add models providers
           git commit -m "$TITLE"
-          git push --force-with-lease origin "$BRANCH"
+          bun sync:auto-merge HEAD^ HEAD
+          safe="$(sed -n 's/^safe=//p' "$GITHUB_OUTPUT" | tail -1)"
 
           pr_number="$(gh pr list --head "$BRANCH" --base dev --json number --jq '.[0].number')"
+          if [ "$safe" != "true" ] && [ -n "$pr_number" ]; then
+            gh pr merge "$pr_number" --disable-auto || true
+            if [ "$(gh pr view "$pr_number" --json autoMergeRequest --jq '.autoMergeRequest == null')" != "true" ]; then
+              echo "Failed to disable auto-merge for unsafe sync PR #$pr_number."
+              exit 1
+            fi
+          fi
+
+          git push --force-with-lease origin "$BRANCH"
+
           if [ -n "$pr_number" ]; then
             gh pr edit "$pr_number" --title "$TITLE" --body-file .sync/model-sync-report.md
             for label in "${labels[@]}"; do
@@ -129,4 +141,12 @@ jobs:
             done
           else
             gh pr create --base dev --head "$BRANCH" --title "$TITLE" --body-file .sync/model-sync-report.md "${label_args[@]}"
+            pr_number="$(gh pr list --head "$BRANCH" --base dev --json number --jq '.[0].number')"
+          fi
+
+          if [ "$safe" = "true" ]; then
+            gh pr merge "$pr_number" --auto --squash
+          elif [ "$(gh pr view "$pr_number" --json autoMergeRequest --jq '.autoMergeRequest == null')" != "true" ]; then
+            echo "Unsafe sync PR #$pr_number still has auto-merge enabled."
+            exit 1
           fi

.opencode/agent/pr-reviewer.md

@@ -12,6 +12,7 @@ permission:
     "*.env.*": deny
   glob: allow
   grep: allow
+  mark-pr-ready: allow
   external_directory: deny
 ---
 
@@ -65,6 +66,8 @@ Focus only on actionable problems introduced by the pull request:
 
 Do not report style preferences, speculative concerns, pre-existing problems, or bare schema errors that validation will identify without useful explanation. Do not invent requirements from neighboring files when provider behavior is intentionally different. Do not claim to have run commands, opened links, or performed validation. Do not edit files or attempt to post comments yourself.
 
+Use `mark-pr-ready` only after completing the review and determining there are no action items. Never use it when returning one or more action items.
+
 Every finding must be an action item: the author must need to change something, verify a specific fact, or provide missing evidence. Do not list checks that passed or general observations. If you find action items, list them in severity order and return exactly this structure:
 
 ```markdown
@@ -74,6 +77,6 @@ Every finding must be an action item: the author must need to change something,
 
 Use `violation` only when the change demonstrably breaks a repository requirement or expected behavior. Use `possible mistake` when the diff provides concrete contradictory or suspicious evidence but external facts must be verified. Use `critical`, `high`, `medium`, or `low` for severity. Reference a changed line whenever possible and keep each action item concise.
 
-If there are no action items, respond with exactly the following text and nothing else. Do not explain what you checked or why it passed:
+If there are no action items, call `mark-pr-ready`, then respond with exactly the following text and nothing else. Do not explain what you checked or why it passed:
 
 `No actionable findings.`

.opencode/opencode.jsonc

@@ -0,0 +1,6 @@
+{
+  "$schema": "https://opencode.ai/config.json",
+  "permission": {
+    "mark-pr-ready": "deny"
+  }
+}

.opencode/tool/mark-pr-ready.ts

@@ -0,0 +1,16 @@
+import { writeFile } from "node:fs/promises"
+import { tool } from "@opencode-ai/plugin"
+
+export default tool({
+  description: "Mark the current pull request as ready after completing a review with no actionable findings.",
+  args: {},
+  async execute(_args, context) {
+    if (context.agent !== "pr-reviewer") throw new Error("This tool is only available to the pr-reviewer agent")
+
+    const readyFile = process.env.PR_REVIEW_READY_FILE
+    if (!readyFile) throw new Error("PR_REVIEW_READY_FILE is not configured")
+
+    await writeFile(readyFile, "")
+    return "Pull request marked ready."
+  },
+})

labs/arcee-ai/lab.toml

@@ -0,0 +1 @@
+description = "Arcee AI develops open-weight language models focused on efficient reasoning, tool use, and deployable intelligence."

labs/arcee-ai/logo.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" fill-rule="evenodd"><path d="M13.236 2.377 2.751 20.493H0L11.863 0l1.373 2.377zm3.554 6.156-9.606 11.96H4.13L15.511 6.32l1.279 2.212zm6.908 11.96H14.05l8.406-2.151 1.242 2.15zm-3.42-5.922-7.843 5.92H8.482l10.597-7.997 1.2 2.077z"/></svg>

models/aisingapore/gemma-sea-lion-v4-27b-it.toml

@@ -0,0 +1,22 @@
+name = "Gemma-SEA-LION-v4-27B-IT"
+description = "Gemma 3 27B tuned by AI Singapore for Southeast Asian languages and instruction following"
+family = "gemma"
+release_date = "2025-09-23"
+last_updated = "2025-09-23"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+open_weights = true
+
+[limit]
+context = 128_000
+output = 128_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/aisingapore/Gemma-SEA-LION-v4-27B-IT"

models/alibaba/qwen2.5-coder-0.5b.toml

@@ -0,0 +1,23 @@
+name = "Qwen2.5-Coder-0.5B"
+description = "Tiny open Qwen code model for lightweight completion and on-device coding"
+family = "qwen"
+release_date = "2024-11-12"
+last_updated = "2024-11-12"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+open_weights = true
+license = "Apache 2.0"
+
+[limit]
+context = 32_768
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B"

models/alibaba/qwen2.5-coder-32b-instruct.toml

@@ -0,0 +1,22 @@
+name = "Qwen2.5-Coder-32B-Instruct"
+description = "Open coding-focused Qwen model for code generation, repair, and repository reasoning"
+family = "qwen"
+release_date = "2024-11-12"
+last_updated = "2024-11-12"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct"

models/alibaba/qwen3-235b-a22b-instruct-2507.toml

@@ -0,0 +1,23 @@
+name = "Qwen3 235B-A22B Instruct 2507"
+description = "Updated large open Qwen3 MoE instruct model for multilingual chat, coding, and tool use"
+family = "qwen"
+release_date = "2025-07-21"
+last_updated = "2025-07-21"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+license = "Apache 2.0"
+
+[limit]
+context = 262_144
+output = 16_384
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507"

models/alibaba/qwen3-30b-a3b.toml

@@ -0,0 +1,22 @@
+name = "Qwen3 30B A3B"
+description = "Sparse MoE Qwen model with 3B active parameters for efficient chat and reasoning"
+family = "qwen"
+release_date = "2025-04-28"
+last_updated = "2025-04-28"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 131_072
+output = 16_384
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-30B-A3B"

models/alibaba/qwen3-coder-next.toml

@@ -0,0 +1,27 @@
+# https://qwen.ai/blog?id=qwen3-coder-next
+# https://huggingface.co/Qwen/Qwen3-Coder-Next
+# https://www.qwencloud.com/models/qwen3-coder-next
+name = "Qwen3 Coder Next"
+description = "Open-weight Qwen coding model for agents, repository edits, and multi-turn tool use"
+family = "qwen"
+release_date = "2026-02-03"
+last_updated = "2026-02-03"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-09"
+open_weights = true
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-Coder-Next"

models/alibaba/qwen3.5-flash.toml

@@ -0,0 +1,22 @@
+# https://help.aliyun.com/en/model-studio/qwen3-5-flash
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+name = "Qwen3.5 Flash"
+description = "Qwen vision-language model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2026-02-23"
+last_updated = "2026-02-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+
+[limit]
+context = 1_000_000
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/alibaba/qwen3.7-flash.toml

@@ -0,0 +1,20 @@
+name = "Qwen3.7 Flash"
+description = "Lightweight multimodal Qwen model for high-throughput text, image, and video tasks"
+family = "qwen"
+release_date = "2026-07-15"
+last_updated = "2026-07-15"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+input = 991_000
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/alibaba/qwen3.8-27b.toml

@@ -0,0 +1,36 @@
+# Sources (accessed 2026-08-15):
+# https://huggingface.co/Qwen/Qwen3.8-27B
+# https://huggingface.co/api/models/Qwen/Qwen3.8-27B
+# https://qwen.ai/blog?id=qwen3.8
+# Hub lastModified 2026-08-14T15:00:01Z is the open-weight drop.
+# Do not use Hub createdAt 2026-08-05 (staged countdown page).
+
+name = "Qwen3.8 27B"
+description = "Dense 27B vision-language model for coding, agent tasks, and image and video understanding"
+family = "qwen"
+release_date = "2026-08-14"
+last_updated = "2026-08-14"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 262_144
+output = 32_768
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.8-27B"
+
+[[benchmarks]]
+name = "SWE-bench Pro"
+score = 61.7
+metric = "resolved"
+source = "https://huggingface.co/Qwen/Qwen3.8-27B"

models/alibaba/qwen3.8-max-preview.toml

@@ -28,3 +28,131 @@ output = 131_072
 [modalities]
 input = ["text", "image", "video"]
 output = ["text"]
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 86.6
+metric = "accuracy"
+variant = "xhigh"
+version = "2.1"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 67.7
+metric = "resolve rate"
+variant = "xhigh"
+harness = "Claude Code"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 56.6
+metric = "resolve rate"
+variant = "xhigh"
+harness = "Claude Code"
+version = "1.1"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 55.9
+metric = "resolve rate"
+variant = "xhigh"
+harness = "Claude Code"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "FrontierSWE"
+score = 73.5
+metric = "dominance score"
+variant = "xhigh"
+harness = "Claude Code"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "MLS-Bench-Lite"
+score = 41.0
+metric = "score"
+variant = "xhigh"
+harness = "Claude Code"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 27.3
+metric = "pass@1"
+variant = "xhigh"
+dataset = "600-task public subset"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "Toolathlon Verified"
+score = 72.5
+metric = "pass@1"
+variant = "xhigh"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "WideSearch"
+score = 81.9
+metric = "F1"
+variant = "xhigh"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 56.2
+metric = "accuracy"
+variant = "xhigh, with tools"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 92.6
+metric = "accuracy"
+variant = "xhigh"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 43.6
+metric = "accuracy"
+variant = "xhigh, no tools"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "IFBench"
+score = 82.8
+metric = "score"
+variant = "xhigh"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 86.1
+metric = "success rate"
+variant = "xhigh"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"
+
+[[benchmarks]]
+name = "MMMU Pro"
+score = 82.3
+metric = "accuracy"
+variant = "xhigh"
+source = "https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421"
+date = "2026-08-03"

models/alibaba/qwen3.8-max.toml

@@ -0,0 +1,38 @@
+# Sources (accessed 2026-08-06):
+# https://www.qwencloud.com/models/qwen3.8-max
+# https://www.qianwenai.com/models/qwen3.8-max
+# https://help.aliyun.com/zh/model-studio/qwen3-8-max
+# https://www.alibabacloud.com/help/en/model-studio/qwen3-8-max
+# https://help.aliyun.com/zh/model-studio/pdf-understanding
+# https://platform.qianwenai.com/docs/developer-guides/tool-calling/pdf-understanding
+# https://docs.qwencloud.com/token-plan/personal/token-plan-personal-overview
+# https://help.aliyun.com/zh/model-studio/token-plan-personal-overview
+# https://help.aliyun.com/en/model-studio/token-plan-personal-overview
+# https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models
+# https://docs.qwencloud.com/developer-guides/text-generation/thinking
+# https://docs.qwencloud.com/developer-guides/clients-and-developer-tools/opencode
+# https://platform.qianwenai.com/docs/developer-guides/clients-and-developer-tools/opencode
+# https://qwen.ai/blog?id=qwen3.8
+# PDF input: Model Studio / 千问AI docs list only qwen3.8-max under PDF理解
+# (type:file / file_url|file_data). Model pages list Image/Text/Video badges
+# and separately list PDF理解 as a Completions built-in tool. Beijing-region
+# availability note on help.aliyun.com; lab capability still includes pdf.
+
+name = "Qwen3.8 Max"
+description = "2.4-trillion-parameter MoE flagship for coding, professional work, multimodal understanding, and long-horizon agentic workflows"
+family = "qwen"
+release_date = "2026-08-03"
+last_updated = "2026-08-03"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text", "image", "video", "pdf"]
+output = ["text"]

models/alibaba/qwq-32b.toml

@@ -0,0 +1,23 @@
+name = "QwQ 32B"
+description = "Open reasoning model from the Qwen team for math, coding, and step-by-step problem solving"
+family = "qwen"
+release_date = "2025-03-05"
+last_updated = "2025-03-05"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+knowledge = "2024-04"
+open_weights = true
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/QwQ-32B"

models/anthropic/claude-mythos-5.toml

@@ -0,0 +1,23 @@
+# Sources:
+# https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5
+# https://www.anthropic.com/claude/mythos
+name = "Claude Mythos 5"
+description = "Restricted Claude model for advanced cybersecurity and biology research workflows"
+family = "claude-mythos"
+release_date = "2026-06-09"
+last_updated = "2026-06-09"
+attachment = true
+reasoning = true
+temperature = false
+tool_call = true
+structured_output = true
+knowledge = "2026-01-31"
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 128_000
+
+[modalities]
+input = ["text", "image", "pdf"]
+output = ["text"]

models/anthropic/claude-opus-5.toml

@@ -17,3 +17,156 @@ output = 128_000
 [modalities]
 input = ["text", "image", "pdf"]
 output = ["text"]
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 96.0
+metric = "resolved"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 79.2
+metric = "resolve rate"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "SWE-Bench Multilingual"
+score = 89.5
+metric = "resolve rate"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "SWE-Bench Multimodal"
+score = 59.4
+metric = "resolve rate"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 68.8
+metric = "resolve rate"
+version = "1.1"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "FrontierCode"
+score = 53.4
+metric = "mean@5"
+variant = "medium effort"
+dataset = "Main"
+version = "1.1"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "Frontier-Bench"
+score = 43.3
+metric = "mean reward"
+variant = "max effort"
+harness = "mini-SWE-agent"
+version = "v0.1"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "BrowseComp"
+score = 90.8
+metric = "accuracy"
+variant = "single agent"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 56.3
+metric = "accuracy"
+variant = "no tools"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 64.7
+metric = "accuracy"
+variant = "with tools"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "DeepSearchQA"
+score = 95.0
+metric = "F1"
+variant = "max effort"
+source = "https://www.anthropic.com/news/claude-opus-5"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "OSWorld"
+score = 70.6
+metric = "success rate"
+version = "2.0"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1861
+metric = "Elo"
+variant = "max effort"
+version = "v2"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "AA-Briefcase"
+score = 1720
+metric = "Elo"
+variant = "max effort"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 26.0
+metric = "success rate"
+variant = "max effort"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "ARC-AGI-1"
+score = 97.5
+metric = "accuracy"
+variant = "max effort"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "ARC-AGI-2"
+score = 90.4
+metric = "accuracy"
+variant = "max effort"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "ARC-AGI-3"
+score = 30.2
+metric = "RHAE"
+variant = "high effort"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"
+
+[[benchmarks]]
+name = "HealthBench Professional"
+score = 59.8
+metric = "score"
+variant = "max effort"
+source = "https://www.anthropic.com/claude-opus-5-system-card"
+date = "2026-07-24"

models/arcee-ai/trinity-large-preview.toml

@@ -0,0 +1,40 @@
+# Source: https://huggingface.co/arcee-ai/Trinity-Large-Preview
+name = "Trinity Large Preview"
+description = "Lightly post-trained 398B MoE chat model for creative work, long-context prompts, and tool-using agents"
+family = "trinity"
+release_date = "2026-01-27"
+last_updated = "2026-05-28"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+license = "OpenMDW-1.1"
+
+[limit]
+context = 524_288
+output = 262_144
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/arcee-ai/Trinity-Large-Preview"
+format = "safetensors"
+
+[[links]]
+label = "Model card"
+url = "https://huggingface.co/arcee-ai/Trinity-Large-Preview"
+type = "model_card"
+
+[[links]]
+label = "Announcement"
+url = "https://www.arcee.ai/blog/trinity-large"
+type = "announcement"
+
+[[links]]
+label = "License"
+url = "https://huggingface.co/arcee-ai/Trinity-Large-Preview/blob/main/LICENSE"
+type = "license"

models/arcee-ai/trinity-large-thinking.toml

@@ -0,0 +1,40 @@
+# Source: https://huggingface.co/arcee-ai/Trinity-Large-Thinking
+name = "Trinity Large Thinking"
+description = "Reasoning-optimized 398B MoE agent model with extended thinking for long-horizon and multi-turn tool use"
+family = "trinity"
+release_date = "2026-04-01"
+last_updated = "2026-05-28"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+license = "OpenMDW-1.1"
+
+[limit]
+context = 524_288
+output = 262_144
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/arcee-ai/Trinity-Large-Thinking"
+format = "safetensors"
+
+[[links]]
+label = "Model card"
+url = "https://huggingface.co/arcee-ai/Trinity-Large-Thinking"
+type = "model_card"
+
+[[links]]
+label = "Announcement"
+url = "https://www.arcee.ai/blog/trinity-large-thinking"
+type = "announcement"
+
+[[links]]
+label = "License"
+url = "https://huggingface.co/arcee-ai/Trinity-Large-Thinking/blob/main/LICENSE"
+type = "license"

models/arcee-ai/trinity-mini.toml

@@ -0,0 +1,40 @@
+# Source: https://huggingface.co/arcee-ai/Trinity-Mini
+name = "Trinity Mini"
+description = "Reasoning-tuned 26B MoE model with 3B active parameters for agents, tools, and multi-step workloads"
+family = "trinity"
+release_date = "2025-12-01"
+last_updated = "2026-05-28"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+license = "OpenMDW-1.1"
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/arcee-ai/Trinity-Mini"
+format = "safetensors"
+
+[[links]]
+label = "Model card"
+url = "https://huggingface.co/arcee-ai/Trinity-Mini"
+type = "model_card"
+
+[[links]]
+label = "Announcement"
+url = "https://www.arcee.ai/blog/the-trinity-manifesto"
+type = "announcement"
+
+[[links]]
+label = "License"
+url = "https://huggingface.co/arcee-ai/Trinity-Mini/blob/main/LICENSE"
+type = "license"

models/arcee-ai/trinity-nano-preview.toml

@@ -0,0 +1,40 @@
+# Source: https://huggingface.co/arcee-ai/Trinity-Nano-Preview
+name = "Trinity Nano Preview"
+description = "Experimental chat-tuned 6B MoE model with 1B active parameters for low-resource chat and instruction following"
+family = "trinity"
+release_date = "2025-12-01"
+last_updated = "2026-05-28"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+license = "OpenMDW-1.1"
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/arcee-ai/Trinity-Nano-Preview"
+format = "safetensors"
+
+[[links]]
+label = "Model card"
+url = "https://huggingface.co/arcee-ai/Trinity-Nano-Preview"
+type = "model_card"
+
+[[links]]
+label = "Announcement"
+url = "https://www.arcee.ai/blog/the-trinity-manifesto"
+type = "announcement"
+
+[[links]]
+label = "License"
+url = "https://huggingface.co/arcee-ai/Trinity-Nano-Preview/blob/main/LICENSE"
+type = "license"

models/bytedance-seed/seed-1-6-flash.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 1.6 Flash"
+description = "Low-latency ByteDance Seed model for high-throughput chat, extraction, and lightweight tool use"
+family = "seed"
+release_date = "2025-08-28"
+last_updated = "2025-08-28"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 32_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/bytedance-seed/seed-1-6-vision.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 1.6 Vision"
+description = "ByteDance Seed multimodal model for image understanding, visual reasoning, and tool-assisted tasks"
+family = "seed"
+release_date = "2025-08-15"
+last_updated = "2025-08-15"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 32_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/bytedance-seed/seed-1-6.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 1.6"
+description = "ByteDance Seed model for long-context reasoning, instruction following, and tool-assisted tasks"
+family = "seed"
+release_date = "2025-10-15"
+last_updated = "2025-10-15"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 64_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/bytedance-seed/seed-1-8.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 1.8"
+description = "ByteDance Seed model for multimodal reasoning, long-context analysis, and agent workflows"
+family = "seed"
+release_date = "2025-12-28"
+last_updated = "2025-12-28"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 64_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/bytedance-seed/seed-2.0-code.toml

@@ -0,0 +1,23 @@
+# Sources (accessed 2026-08-11):
+# - https://seed.bytedance.com/en/blog/seed-2-0-official-launch
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 2.0 Code"
+description = "ByteDance Seed coding model for multimodal software engineering and long-running agents"
+family = "seed"
+release_date = "2026-02-14"
+last_updated = "2026-02-14"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 262_144
+output = 131_072
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-2.0-lite.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 2.0 Lite"
+description = "Cost-efficient ByteDance Seed 2.0 model for production chat, analysis, and structured generation"
+family = "seed"
+release_date = "2026-02-14"
+last_updated = "2026-02-14"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 32_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-2.0-mini.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 2.0 Mini"
+description = "Lightweight ByteDance Seed 2.0 model for low-latency multimodal reasoning and high-volume tasks"
+family = "seed"
+release_date = "2026-02-14"
+last_updated = "2026-02-14"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 32_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-2.0-pro.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 2.0 Pro"
+description = "Flagship ByteDance Seed 2.0 model for complex multimodal reasoning and long-horizon agent workflows"
+family = "seed"
+release_date = "2026-02-14"
+last_updated = "2026-02-14"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 128_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-2.1-pro.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 2.1 Pro"
+description = "Flagship ByteDance Seed 2.1 model for complex multimodal reasoning, coding, and agents"
+family = "seed"
+release_date = "2026-06-23"
+last_updated = "2026-06-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 256_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-2.1-turbo.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed 2.1 Turbo"
+description = "Faster ByteDance Seed 2.1 model for multimodal reasoning and latency-sensitive agent workflows"
+family = "seed"
+release_date = "2026-06-23"
+last_updated = "2026-06-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 256_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-character.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed Character"
+description = "ByteDance Seed model optimized for character-driven dialogue and consistent conversational behavior"
+family = "seed"
+release_date = "2026-06-23"
+last_updated = "2026-06-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 256_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/bytedance-seed/seed-evolving.toml

@@ -0,0 +1,21 @@
+# Sources (accessed 2026-08-14):
+# - https://seed.bytedance.com/en/seed2
+# - https://www.volcengine.com/docs/82379/1330310
+name = "Seed Evolving"
+description = "Rolling ByteDance Seed model for rapidly updated reasoning, coding, and agent capabilities"
+family = "seed"
+release_date = "2026-06-23"
+last_updated = "2026-06-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 256_000
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]

models/deepseek/deepseek-ocr-2.toml

@@ -0,0 +1,16 @@
+name = "DeepSeek OCR 2"
+description = "High-accuracy OCR model for extracting text from documents, screenshots, receipts, and natural scenes"
+release_date = "2026-01-27"
+last_updated = "2026-01-27"
+attachment = true
+reasoning = false
+tool_call = false
+open_weights = true
+
+[limit]
+context = 8_192
+output = 8_192
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/deepseek/deepseek-r1-distill-qwen-32b.toml

@@ -0,0 +1,22 @@
+name = "DeepSeek-R1-Distill-Qwen-32B"
+description = "R1 reasoning distilled into Qwen 2.5 32B for efficient open-weight step-by-step problem solving"
+family = "deepseek-thinking"
+release_date = "2025-01-20"
+last_updated = "2025-01-20"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = false
+open_weights = true
+
+[limit]
+context = 131_072
+output = 32_768
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B"

models/deepseek/deepseek-v3-0324.toml

@@ -0,0 +1,23 @@
+name = "DeepSeek V3 0324"
+description = "March 2025 checkpoint of DeepSeek-V3 with improved reasoning and coding"
+family = "deepseek"
+release_date = "2025-03-24"
+last_updated = "2025-03-24"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 163_840
+output = 163_840
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Model weights"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V3-0324"
+format = "safetensors"

models/deepseek/deepseek-v3.1.toml

@@ -0,0 +1,23 @@
+name = "DeepSeek-V3.1"
+description = "Hybrid-reasoning DeepSeek model with thinking and non-thinking modes"
+family = "deepseek"
+release_date = "2025-08-21"
+last_updated = "2025-08-21"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+license = "MIT License"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V3.1"

models/deepseek/deepseek-v3.2.toml

@@ -0,0 +1,27 @@
+# https://api-docs.deepseek.com/news/news251201
+# https://huggingface.co/deepseek-ai/DeepSeek-V3.2
+name = "DeepSeek V3.2"
+description = "Hybrid-reasoning DeepSeek model with thinking and non-thinking modes, sparse attention, and tool-use"
+family = "deepseek"
+release_date = "2025-12-01"
+last_updated = "2025-12-01"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2024-07"
+open_weights = true
+license = "MIT License"
+
+[limit]
+context = 128_000
+output = 64_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V3.2"

models/deepseek/deepseek-v3.toml

@@ -0,0 +1,23 @@
+name = "DeepSeek-V3"
+description = "Open DeepSeek MoE chat model for coding, math, and general reasoning"
+family = "deepseek"
+release_date = "2024-12-26"
+last_updated = "2024-12-26"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+license = "DeepSeek Model License"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V3"

models/deepseek/deepseek-v4-flash-0731.toml

@@ -0,0 +1,105 @@
+name = "DeepSeek V4 Flash 0731"
+description = "Official DeepSeek V4 Flash release with enhanced agentic capabilities and integrated DSpark speculative decoding"
+family = "deepseek-flash"
+release_date = "2026-07-31"
+last_updated = "2026-07-31"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-05"
+open_weights = true
+license = "MIT"
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 82.7
+metric = "pass@1"
+variant = "max"
+version = "2.1"
+source = "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 54.2
+metric = "resolve rate"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "CyberGym"
+score = 76.7
+metric = "score"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 54.4
+metric = "resolve rate"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "Toolathlon-Verified"
+score = 70.3
+metric = "score"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "Agents' Last Exam"
+score = 25.2
+metric = "score"
+variant = "max effort"
+harness = "DeepSeek Harness minimal mode"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 25.1
+metric = "success rate"
+variant = "max effort"
+dataset = "public"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "DSBench-FullStack"
+score = 68.7
+metric = "score"
+variant = "max effort"
+dataset = "internal"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"
+
+[[benchmarks]]
+name = "DSBench-Hard"
+score = 59.6
+metric = "score"
+variant = "max effort"
+dataset = "internal"
+source = "https://api-docs.deepseek.com/updates/"
+date = "2026-07-31"

models/deepseek/deepseek-v4-pro-0813.toml

@@ -0,0 +1,19 @@
+name = "DeepSeek V4 Pro 0813"
+description = "DeepSeek V4 Pro snapshot with million-token context and support for thinking and non-thinking modes"
+family = "deepseek-thinking"
+release_date = "2026-08-12"
+last_updated = "2026-08-12"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/google/gemini-3.5-flash-lite.toml

@@ -17,4 +17,76 @@ output = 65_536
 
 [modalities]
 input = ["text", "image", "video", "audio", "pdf"]
-output = ["text"]
+output = ["text"]
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 54.2
+metric = "resolve rate"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 54.0
+metric = "accuracy"
+harness = "Terminus 2"
+version = "2.1"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "MLE-Bench"
+score = 39.2
+metric = "average position score"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1140
+metric = "Elo"
+version = "v2"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 74.0
+metric = "success rate"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 74.5
+metric = "accuracy"
+variant = "no tools"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 76.5
+metric = "accuracy"
+variant = "with tools"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "GDM-MRCR"
+score = 72.2
+metric = "accuracy"
+variant = "128k average, 8-needle"
+version = "v2"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "GDM-MRCR"
+score = 21.3
+metric = "accuracy"
+variant = "1M pointwise, 8-needle"
+version = "v2"
+source = "https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/"
+date = "2026-07-21"

models/google/gemini-3.6-flash.toml

@@ -17,4 +17,87 @@ output = 65_536
 
 [modalities]
 input = ["text", "image", "video", "audio", "pdf"]
-output = ["text"]
+output = ["text"]
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 58.7
+metric = "resolve rate"
+harness = "Antigravity"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 49.0
+metric = "resolve rate"
+variant = "high reasoning"
+version = "1.1"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 78.0
+metric = "accuracy"
+harness = "Terminus 2"
+version = "2.1"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "MLE-Bench"
+score = 63.9
+metric = "average position score"
+dataset = "Partial 30"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1421
+metric = "Elo"
+version = "v2"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 83.0
+metric = "success rate"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 85.2
+metric = "accuracy"
+variant = "no tools"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 89.4
+metric = "accuracy"
+variant = "with tools"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "GDM-MRCR"
+score = 91.8
+metric = "accuracy"
+variant = "128k average, 8-needle"
+version = "v2"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"
+
+[[benchmarks]]
+name = "GDM-MRCR"
+score = 54.0
+metric = "accuracy"
+variant = "1M pointwise, 8-needle"
+version = "v2"
+source = "https://deepmind.google/models/evals-methodology/gemini-3-6-flash/"
+date = "2026-07-21"

models/google/gemini-3.7-flash.toml

@@ -0,0 +1,68 @@
+name = "Gemini 3.7 Flash"
+description = "High-efficiency Gemini model for agentic workflows, coding, and multimodal reasoning"
+family = "gemini-flash"
+release_date = "2026-08-13"
+last_updated = "2026-08-13"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2026-03"
+open_weights = false
+
+[limit]
+context = 1_048_576
+output = 65_536
+
+[modalities]
+input = ["text", "image", "video", "audio", "pdf"]
+output = ["text"]
+
+[[benchmarks]]
+name = "FrontierCode"
+score = 43.6
+metric = "score"
+version = "1.1 Main"
+source = "https://deepmind.google/models/model-cards/gemini-3-7-flash/"
+date = "2026-08-13"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 65.3
+metric = "resolve rate"
+version = "1.1"
+source = "https://deepmind.google/models/model-cards/gemini-3-7-flash/"
+date = "2026-08-13"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 85.8
+metric = "accuracy"
+version = "2.1"
+source = "https://deepmind.google/models/model-cards/gemini-3-7-flash/"
+date = "2026-08-13"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 30.4
+metric = "accuracy"
+dataset = "private set"
+source = "https://deepmind.google/models/model-cards/gemini-3-7-flash/"
+date = "2026-08-13"
+
+[[benchmarks]]
+name = "GDP.pdf"
+score = 34.0
+metric = "accuracy"
+source = "https://deepmind.google/models/model-cards/gemini-3-7-flash/"
+date = "2026-08-13"
+
+[[benchmarks]]
+name = "GDM-MRCR"
+score = 97.0
+metric = "accuracy"
+variant = "128k average, 8-needle"
+version = "v2"
+source = "https://deepmind.google/models/model-cards/gemini-3-7-flash/"
+date = "2026-08-13"

models/ibm/granite-4-h-micro.toml

@@ -0,0 +1,23 @@
+name = "Granite-4.0-H-Micro"
+description = "Compact open-weight hybrid Granite model for lightweight enterprise chat and tool calling"
+family = "granite"
+release_date = "2025-10-02"
+last_updated = "2025-10-02"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/ibm-granite/granite-4.0-h-micro"

models/ibm/granite-4-h-small.toml

@@ -0,0 +1,23 @@
+name = "Granite-4.0-H-Small"
+description = "Open-weight hybrid model for enterprise chat, coding, retrieval-augmented generation, and tool-calling workloads"
+family = "granite"
+release_date = "2025-10-02"
+last_updated = "2025-10-02"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/ibm-granite/granite-4.0-h-small"

models/meta/llama-3.1-8b-instruct.toml

@@ -0,0 +1,23 @@
+name = "Llama-3.1-8B-Instruct"
+description = "Compact open Llama model for lightweight chat, drafting, and self-hosting"
+family = "llama"
+release_date = "2024-07-23"
+last_updated = "2024-07-23"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+knowledge = "2023-12"
+open_weights = true
+
+[limit]
+context = 128_000
+output = 4_096
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct"

models/meta/llama-3.2-11b-vision-instruct.toml

@@ -0,0 +1,23 @@
+name = "Llama-3.2-11B-Vision-Instruct"
+description = "Open multimodal Llama model for image understanding, captioning, and visual QA"
+family = "llama"
+release_date = "2024-09-25"
+last_updated = "2024-09-25"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+knowledge = "2023-12"
+open_weights = true
+
+[limit]
+context = 128_000
+output = 4_096
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct"

models/meta/llama-3.2-1b.toml

@@ -0,0 +1,24 @@
+name = "Llama-3.2-1B"
+description = "Compact open Llama base model for lightweight and on-device use"
+family = "llama"
+release_date = "2024-09-25"
+last_updated = "2024-09-25"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+knowledge = "2023-12"
+open_weights = true
+license = "Llama 3.2 Community License"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/meta-llama/Llama-3.2-1B"

models/meta/llama-3.2-3b.toml

@@ -0,0 +1,24 @@
+name = "Llama-3.2-3B"
+description = "Small open Llama base model for lightweight text generation and self-hosting"
+family = "llama"
+release_date = "2024-09-25"
+last_updated = "2024-09-25"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+knowledge = "2023-12"
+open_weights = true
+license = "Llama 3.2 Community License"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/meta-llama/Llama-3.2-3B"

models/meta/llama-guard-3-8b.toml

@@ -0,0 +1,23 @@
+name = "Llama-Guard-3-8B"
+description = "Llama 3.1-based safety classifier for moderating prompts and model responses"
+family = "llama"
+release_date = "2024-07-23"
+last_updated = "2024-07-23"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+knowledge = "2023-12"
+open_weights = true
+
+[limit]
+context = 128_000
+output = 4_096
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/meta-llama/Llama-Guard-3-8B"

models/meta/muse-glimmer-30b.toml

@@ -0,0 +1,111 @@
+# Sources:
+# https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
+# https://huggingface.co/meta-models/Muse-Glimmer-30B
+# https://developer.meta.com/ai/models/muse-glimmer/
+
+name = "Muse Glimmer 30B"
+description = "Muse Glimmer is a 30-billion-parameter open-weight multimodal model from Meta Superintelligence Labs, distilled from Muse Spark for always-on local agents, tool use, coding, and image understanding."
+family = "muse"
+release_date = "2026-08-10"
+last_updated = "2026-08-10"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2026-01-04"
+open_weights = true
+license = "Apache 2.0"
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+
+[[links]]
+label = "Announcement"
+url = "https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model"
+type = "announcement"
+
+[[links]]
+label = "Model card"
+url = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+type = "model_card"
+
+[[links]]
+label = "Developer docs"
+url = "https://developer.meta.com/ai/models/muse-glimmer/"
+type = "docs"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 75.5
+metric = "success rate"
+variant = "public"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "DeepSearch QA"
+score = 74.6
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 51.2
+metric = "resolve rate"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "SWE-Bench Verified"
+score = 76.0
+metric = "resolve rate"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 51.7
+metric = "success rate"
+version = "2.1"
+variant = "with terminus2"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 65.9
+metric = "success rate"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "AIME 2026"
+score = 94.7
+metric = "accuracy"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 83.5
+metric = "accuracy"
+variant = "AA"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 78.8
+metric = "accuracy"
+source = "https://huggingface.co/meta-models/Muse-Glimmer-30B"
+date = "2026-08-10"

models/meta/muse-spark-1.2.toml

@@ -0,0 +1,23 @@
+# Sources:
+# https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
+# https://dev.meta.ai/docs/getting-started/models
+
+name = "Muse Spark 1.2"
+description = "Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1 with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows."
+family = "muse"
+release_date = "2026-08-05"
+last_updated = "2026-08-05"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 1_048_576
+output = 131_072
+
+[modalities]
+input = ["text", "image", "video", "pdf", "audio"]
+output = ["text"]

models/microsoft/mai-code-1.1-flash.toml

@@ -0,0 +1,23 @@
+name = "MAI-Code-1.1-Flash"
+description = "Microsoft coding model with native vision support, optimized for fast and efficient software development"
+family = "mai"
+release_date = "2026-08-11"
+last_updated = "2026-08-11"
+attachment = true
+reasoning = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 256_000
+output = 128_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[links]]
+label = "Announcement"
+url = "https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/"
+type = "announcement"

models/microsoft/phi-4-mini.toml

@@ -0,0 +1,30 @@
+name = "Phi-4-mini"
+description = "Compact Microsoft instruction model tuned for efficient coding assistance, reasoning, and low-latency agent tasks"
+family = "phi"
+release_date = "2024-12-11"
+last_updated = "2024-12-11"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+knowledge = "2023-10"
+open_weights = true
+
+[limit]
+context = 128_000
+output = 4_096
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[links]]
+label = "Weights"
+url = "https://huggingface.co/microsoft/Phi-4-mini-instruct"
+type = "weights"
+
+[[benchmarks]]
+name = "MMLU"
+score = 67.3
+metric = "accuracy"
+source = "https://huggingface.co/microsoft/Phi-4-mini-instruct/resolve/main/README.md"

models/minimax/MiniMax-M2-Her.toml

@@ -0,0 +1,19 @@
+# Source: https://api.ofox.ai/v2/models/catalog?include=provider_price&limit=500
+name = "MiniMax-M2 Her"
+description = "MiniMax M2 variant tuned for conversational and character-driven agent interactions"
+family = "minimax"
+release_date = "2026-01-23"
+last_updated = "2026-01-23"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = false
+
+[limit]
+context = 200_000
+output = 131_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/mistral/codestral-22b-v0.1.toml

@@ -0,0 +1,23 @@
+name = "Codestral-22B-v0.1"
+description = "Open Mistral code model for fill-in-the-middle and 80+ programming languages"
+family = "codestral"
+release_date = "2024-05-29"
+last_updated = "2024-05-29"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+open_weights = true
+license = "Mistral AI Non-Production License"
+
+[limit]
+context = 32_768
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/mistralai/Codestral-22B-v0.1"

models/mistral/magistral-small-2506.toml

@@ -0,0 +1,23 @@
+name = "Magistral Small"
+description = "Open Mistral reasoning model for transparent step-by-step problem solving"
+family = "magistral"
+release_date = "2025-06-10"
+last_updated = "2025-06-10"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+license = "Apache 2.0"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/mistralai/Magistral-Small-2506"

models/mistral/ministral-8b-instruct-2410.toml

@@ -0,0 +1,23 @@
+name = "Ministral 8B Instruct"
+description = "Efficient open Mistral edge model for on-device chat and function calling"
+family = "ministral"
+release_date = "2024-10-16"
+last_updated = "2024-10-16"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+license = "Mistral Research License"
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/mistralai/Ministral-8B-Instruct-2410"

models/mistral/mistral-medium-2604.toml

@@ -27,3 +27,11 @@ name = "SWE-Bench Verified"
 score = 77.6
 metric = "resolved"
 source = "https://huggingface.co/mistralai/Mistral-Medium-3.5-128B"
+
+[[benchmarks]]
+name = "τ³-Telecom"
+score = 91.4
+metric = "accuracy"
+variant = "public preview"
+source = "https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/"
+date = "2026-05-22"

models/mistral/mistral-small-3-1-24b-instruct-2503.toml

@@ -0,0 +1,24 @@
+name = "Mistral Small 3.1 24B"
+description = "Efficient multimodal model for instruction following, coding, reasoning, and function calling"
+family = "mistral-small"
+release_date = "2025-03-17"
+last_updated = "2025-03-17"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2024-06"
+open_weights = true
+
+[limit]
+context = 128_000
+output = 16_384
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503"

models/moonshotai/kimi-k3.toml

@@ -17,3 +17,119 @@ output = 131_072
 [modalities]
 input = ["text", "image", "video"]
 output = ["text"]
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 67.5
+metric = "resolve rate"
+variant = "max effort"
+harness = "Kimi Code"
+version = "1.1"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 88.3
+metric = "accuracy"
+variant = "max effort"
+harness = "Kimi Code"
+version = "2.1"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "FrontierSWE"
+score = 81.2
+metric = "dominance score"
+variant = "max effort"
+harness = "Kimi Code"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "Program Bench"
+score = 77.8
+metric = "score"
+variant = "max effort"
+harness = "Kimi Code"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "SWE Marathon"
+score = 42.0
+metric = "resolve rate"
+variant = "max effort"
+harness = "Claude Code"
+version = "1.1"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "GDPval-AA"
+score = 1668
+metric = "Elo"
+variant = "max effort"
+version = "v2"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "AA-Briefcase"
+score = 1548
+metric = "Elo"
+variant = "max effort"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "AutomationBench"
+score = 30.8
+metric = "success rate"
+variant = "max effort"
+dataset = "600-task public subset"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "JobBench"
+score = 52.9
+metric = "score"
+variant = "max effort"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "SpreadsheetBench"
+score = 34.8
+metric = "score"
+variant = "max effort"
+harness = "Claude Code"
+version = "2"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "BrowseComp"
+score = 91.2
+metric = "accuracy"
+variant = "max effort, context compaction"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "CharXiv Reasoning"
+score = 91.3
+metric = "accuracy"
+variant = "max effort, with tools"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"
+
+[[benchmarks]]
+name = "ZeroBench"
+score = 41.0
+metric = "pass@5"
+variant = "max effort, with tools"
+source = "https://www.kimi.com/blog/kimi-k3"
+date = "2026-07-16"

models/nvidia/nemotron-3.5-lightning.toml

@@ -0,0 +1,19 @@
+name = "Nemotron 3.5 Lightning 30B A3B"
+description = "Fast NVIDIA Nemotron MoE for reliable agentic tasks across enterprise workloads"
+family = "nemotron"
+release_date = "2026-08-11"
+last_updated = "2026-08-11"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+
+[limit]
+context = 262_144
+output = 262_144
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/openai/gpt-5.3-codex-spark.toml

@@ -1,25 +1,20 @@
-name = "GPT-5.3-Codex"
+name = "GPT-5.3 Codex Spark"
 description = "Coding-optimized GPT model for repository edits, reviews, and agentic software work"
-family = "gpt-codex"
-release_date = "2026-02-24"
-last_updated = "2026-02-24"
+family = "gpt-codex-spark"
+release_date = "2026-02-05"
+last_updated = "2026-02-05"
 attachment = true
 reasoning = true
-reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "max"] }, { type = "budget_tokens" }]
 temperature = false
 knowledge = "2025-08-31"
 tool_call = true
 structured_output = true
 open_weights = false
 
-[cost]
-input = 1.75
-output = 14.00
-cache_read = 0.175
-
 [limit]
-context = 400_000
-output = 128_000
+context = 128_000
+input = 100_000
+output = 32_000
 
 [modalities]
 input = ["text", "image", "pdf"]

models/openai/gpt-5.4-mini.toml

@@ -19,3 +19,135 @@ output = 128_000
 [modalities]
 input = ["text", "image"]
 output = ["text"]
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 54.4
+metric = "resolve rate"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 60.0
+metric = "accuracy"
+variant = "reasoning effort xhigh"
+version = "2.0"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 57.7
+metric = "score"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Toolathlon"
+score = 42.9
+metric = "score"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "τ²-Bench Telecom"
+score = 93.4
+metric = "accuracy"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 88.0
+metric = "accuracy"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 41.5
+metric = "accuracy"
+variant = "with tools"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 28.2
+metric = "accuracy"
+variant = "without tools"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 72.1
+metric = "success rate"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "MMMU Pro"
+score = 78.0
+metric = "accuracy"
+variant = "with Python"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "MMMU Pro"
+score = 76.6
+metric = "accuracy"
+variant = "without tools"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OmniDocBench"
+score = 0.1263
+metric = "overall edit distance"
+variant = "reasoning effort none"
+version = "1.5"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OpenAI MRCR"
+score = 47.7
+metric = "accuracy"
+variant = "8-needle, 64K-128K"
+version = "v2"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OpenAI MRCR"
+score = 33.6
+metric = "accuracy"
+variant = "8-needle, 128K-256K"
+version = "v2"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Graphwalks"
+score = 76.3
+metric = "accuracy"
+variant = "BFS, 0-128K"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Graphwalks"
+score = 71.5
+metric = "accuracy"
+variant = "parents, 0-128K"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"

models/openai/gpt-5.4-nano.toml

@@ -19,3 +19,135 @@ output = 128_000
 [modalities]
 input = ["text", "image"]
 output = ["text"]
+
+[[benchmarks]]
+name = "SWE-Bench Pro"
+score = 52.4
+metric = "resolve rate"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 46.3
+metric = "accuracy"
+variant = "reasoning effort xhigh"
+version = "2.0"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 56.1
+metric = "score"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Toolathlon"
+score = 35.5
+metric = "score"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "τ²-Bench Telecom"
+score = 92.5
+metric = "accuracy"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 82.8
+metric = "accuracy"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 37.7
+metric = "accuracy"
+variant = "with tools"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 24.3
+metric = "accuracy"
+variant = "without tools"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OSWorld-Verified"
+score = 39.0
+metric = "success rate"
+variant = "reasoning effort xhigh"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "MMMU Pro"
+score = 69.5
+metric = "accuracy"
+variant = "with Python"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "MMMU Pro"
+score = 66.1
+metric = "accuracy"
+variant = "without tools"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OmniDocBench"
+score = 0.2419
+metric = "overall edit distance"
+variant = "reasoning effort none"
+version = "1.5"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OpenAI MRCR"
+score = 44.2
+metric = "accuracy"
+variant = "8-needle, 64K-128K"
+version = "v2"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "OpenAI MRCR"
+score = 33.1
+metric = "accuracy"
+variant = "8-needle, 128K-256K"
+version = "v2"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Graphwalks"
+score = 73.4
+metric = "accuracy"
+variant = "BFS, 0-128K"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"
+
+[[benchmarks]]
+name = "Graphwalks"
+score = 50.8
+metric = "accuracy"
+variant = "parents, 0-128K"
+source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
+date = "2026-03-17"

models/sdaia/allam-2-7b.toml

@@ -0,0 +1,21 @@
+name = "ALLaM-2-7b"
+description = "ALLaM-2-7b instruction tuned model by SDAIA"
+release_date = "2025-01-23"
+last_updated = "2025-01-23"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = false
+open_weights = true
+
+[limit]
+context = 4096
+output = 4096
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/ALLaM-AI/ALLaM-2.0-7B-Instruct"

models/swiss-ai/apertus-70b.toml

@@ -0,0 +1,28 @@
+name = "Apertus 70B"
+description = "Fully open 70B multilingual LLM supporting 1800+ languages with 65K context. Trained on 15T tokens of compliant open data. Apache 2.0, EU AI Act compliant."
+release_date = "2025-09-02"
+last_updated = "2025-09-02"
+knowledge = "2025-09"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+license = "Apache-2.0"
+
+[limit]
+context = 65_536
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509"
+
+[[links]]
+label = "Paper"
+url = "https://arxiv.org/abs/2509.14233"
+type = "paper"

models/xai/grok-4.1-fast.toml

@@ -0,0 +1,24 @@
+# xAI Grok 4.1 Fast (non-reasoning).
+# Sources:
+# - https://x.ai/news/grok-4-1-fast (release 2025-11-19; variants + $0.20/$0.50/$0.05 pricing)
+# - https://docs.oracle.com/en-us/iaas/Content/generative-ai/xai-grok-4-1-fast.htm (2M context, text+image, tools, structured outputs, non-reasoning mode)
+# - https://api.ofox.ai/v1/models/x-ai/grok-4.1-fast (canonical_slug grok-4-1-fast-non-reasoning; context 2M; max_completion 30k)
+name = "Grok 4.1 Fast"
+description = "xAI's fast agentic tool-calling model with a 2M context window; non-reasoning variant for low-latency responses"
+family = "grok"
+release_date = "2025-11-19"
+last_updated = "2025-11-19"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 2_000_000
+output = 30_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/xai/grok-4.5.toml

@@ -1,5 +1,5 @@
 name = "Grok 4.5"
-description = "xAI's latest Grok for chat, coding, agentic tools, and lower hallucination risk"
+description = "xAI's Grok model for chat, coding, agentic tools, and lower hallucination risk"
 family = "grok"
 release_date = "2026-07-08"
 last_updated = "2026-07-08"
@@ -56,3 +56,10 @@ harness = "mini-swe-agent"
 version = "1.1"
 source = "https://x.ai/news/grok-4-5"
 date = "2026-07-08"
+
+[[benchmarks]]
+name = "SWE Marathon"
+score = 29.0
+metric = "pass@1"
+source = "https://x.ai/news/grok-4-5"
+date = "2026-07-08"

models/xai/grok-4.6.toml

@@ -0,0 +1,20 @@
+name = "Grok 4.6"
+description = "xAI's frontier model for long-running agents, coding, knowledge work, and visual projects"
+family = "grok"
+knowledge = "2026-02-01"
+release_date = "2026-08-12"
+last_updated = "2026-08-12"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 500_000
+output = 500_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/zhipuai/glm-5.2.toml

@@ -44,3 +44,124 @@ score = 74.4
 metric = "dominance"
 source = "https://z.ai/blog/glm-5.2"
 date = "2026-06-16"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 40.5
+metric = "accuracy"
+dataset = "text-only subset"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Humanity's Last Exam"
+score = 54.7
+metric = "accuracy"
+variant = "with tools"
+dataset = "text-only subset"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "CritPt"
+score = 20.9
+metric = "accuracy"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "AIME"
+score = 99.2
+metric = "accuracy"
+version = "2026"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "HMMT"
+score = 94.4
+metric = "accuracy"
+version = "November 2025"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "HMMT"
+score = 92.5
+metric = "accuracy"
+version = "February 2026"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "IMOAnswerBench"
+score = 91.0
+metric = "accuracy"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "GPQA Diamond"
+score = 91.2
+metric = "accuracy"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "NL2Repo"
+score = 48.9
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "DeepSWE"
+score = 46.2
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Program Bench"
+score = 63.7
+metric = "score"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Terminal-Bench"
+score = 81.0
+metric = "success rate"
+harness = "Terminus 2"
+version = "2.1"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "PostTrainBench"
+score = 34.3
+metric = "score"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "SWE Marathon"
+score = 13.0
+metric = "resolve rate"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "MCP Atlas"
+score = 76.8
+metric = "score"
+dataset = "public subset"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"
+
+[[benchmarks]]
+name = "Tool-Decathlon"
+score = 48.2
+metric = "score"
+source = "https://z.ai/blog/glm-5.2"
+date = "2026-06-16"

models/zhipuai/glm-5.3.toml

@@ -0,0 +1,19 @@
+name = "GLM-5.3"
+description = "Flagship GLM model for long-horizon coding, agents, and complex project delivery"
+family = "glm"
+release_date = "2026-08-14"
+last_updated = "2026-08-14"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]

package.json

@@ -28,6 +28,7 @@
     "huggingface:sync": "bun ./packages/core/script/sync-models.ts huggingface",
     "kilo:sync": "bun ./packages/core/script/sync-models.ts kilo",
     "llmgateway:sync": "bun ./packages/core/script/sync-models.ts llmgateway",
+    "requesty:sync": "bun ./packages/core/script/sync-models.ts requesty",
     "merge-gateway:sync": "bun ./packages/core/script/sync-models.ts merge-gateway",
     "nano-gpt:sync": "bun ./packages/core/script/sync-models.ts nano-gpt",
     "venice:sync": "bun ./packages/core/script/sync-models.ts venice",
@@ -37,7 +38,8 @@
     "digitalocean:sync": "bun ./packages/core/script/sync-models.ts digitalocean",
     "ambient:sync": "bun ./packages/core/script/sync-models.ts ambient",
     "models:sync": "bun ./packages/core/script/sync-models.ts",
-    "sync:models": "bun ./packages/core/script/sync-models.ts"
+    "sync:models": "bun ./packages/core/script/sync-models.ts",
+    "sync:auto-merge": "bun ./packages/core/script/check-sync-auto-merge.ts"
   },
   "dependencies": {
     "@cloudflare/workers-types": "^4.20260424.1",

packages/core/script/check-sync-auto-merge.ts

@@ -0,0 +1,31 @@
+import { appendFile } from "node:fs/promises";
+
+import { classifyAutoMerge, parseNameStatus } from "../src/sync/auto-merge.js";
+
+const base = process.argv[2] ?? "HEAD^";
+const head = process.argv[3] ?? "HEAD";
+const diff = Bun.spawnSync(["git", "diff", "--name-status", "--no-renames", base, head], {
+  stdout: "pipe",
+  stderr: "inherit",
+});
+
+if (diff.exitCode !== 0) process.exit(diff.exitCode ?? 1);
+
+const loadPrevious = async (path: string) => {
+  const file = Bun.spawnSync(["git", "show", `${base}:${path}`], {
+    stdout: "pipe",
+    stderr: "inherit",
+  });
+  if (file.exitCode !== 0) throw new Error(`Failed to read ${path} at ${base}`);
+  return file.stdout.toString();
+};
+
+const decision = await classifyAutoMerge(parseNameStatus(diff.stdout.toString()), undefined, loadPrevious);
+const summary = decision.safe
+  ? `Safe to auto-merge: ${decision.created} created, ${decision.updated} updated, ${decision.deleted} deleted.`
+  : `Manual review required: ${decision.reasons.join("; ")}.`;
+
+console.log(summary);
+if (process.env.GITHUB_OUTPUT) {
+  await appendFile(process.env.GITHUB_OUTPUT, `safe=${decision.safe}\nsummary=${summary}\n`);
+}

packages/core/src/family.ts

@@ -30,6 +30,7 @@ export const ModelFamilyValues = [
   "claude-sonnet",
   "claude-opus",
   "claude-fable",
+  "claude-mythos",
 
   // Gemini style
   "gemini",
@@ -58,6 +59,7 @@ export const ModelFamilyValues = [
   "qwen3.6",
   "qwen3.7-plus",
   "qwen3.7-max",
+  "qwen3.8-max",
   "qwen-free",
 
   // DeepReinforce

packages/core/src/generate.ts

@@ -89,6 +89,11 @@ async function generateProviders(
     }
 
     const modelsPath = path.join(directory, providerID, "models");
+    if (!existsSync(modelsPath)) {
+      throw new Error(`Provider "${providerID}" has no models`, {
+        cause: { providerPath },
+      });
+    }
     for await (const modelPath of new Bun.Glob("**/*.toml").scan({
       cwd: modelsPath,
       absolute: true,
@@ -124,6 +129,11 @@ async function generateProviders(
       }
       provider.data.models[modelID] = normalizeModelCost(model.data);
     }
+    if (Object.keys(provider.data.models).length === 0) {
+      throw new Error(`Provider "${providerID}" has no models`, {
+        cause: { providerPath },
+      });
+    }
     result[providerID] = provider.data;
   }
 

packages/core/src/schema.ts

@@ -397,6 +397,7 @@ export const Provider = z
       const isOpenAI = data.npm === "@ai-sdk/openai";
       const isOpenAIcompatible = data.npm === "@ai-sdk/openai-compatible";
       const isOpenrouter = data.npm === "@openrouter/ai-sdk-provider";
+      const isMergeGateway = data.npm === "merge-gateway-ai-sdk-provider";
       const isAnthropic = data.npm === "@ai-sdk/anthropic";
       const isKiro = data.npm === "kiro-acp-ai-provider";
       const hasApi = data.api !== undefined;
@@ -406,6 +407,8 @@ export const Provider = z
         (isOpenAIcompatible && hasApi) ||
         // openrouter: must have api
         (isOpenrouter && hasApi) ||
+        // Merge Gateway: native provider with an OpenAI-compatible fallback
+        (isMergeGateway && hasApi) ||
         // anthropic: api optional (always allowed)
         isAnthropic ||
         // openai: api optional (always allowed)
@@ -416,6 +419,7 @@ export const Provider = z
         (!isOpenAI &&
           !isOpenAIcompatible &&
           !isOpenrouter &&
+          !isMergeGateway &&
           !isAnthropic &&
           !isKiro &&
           !hasApi)
@@ -423,7 +427,7 @@ export const Provider = z
     },
     {
       message:
-        "'api' is required for openai-compatible and openrouter, optional for anthropic, openai, and kiro, forbidden otherwise",
+        "'api' is required for openai-compatible, openrouter, and Merge Gateway; optional for anthropic, openai, and kiro; forbidden otherwise",
       path: ["api"],
     },
   );

packages/core/src/sync/auto-merge.ts

@@ -0,0 +1,110 @@
+import { readFile } from "node:fs/promises";
+import { isDeepStrictEqual } from "node:util";
+
+export const MAX_CREATED_MODELS = 10;
+export const MAX_DELETED_MODELS = 10;
+export const MAX_MODEL_CHURN = 15;
+const REVIEWED_REASONING_PROVIDERS = new Set([
+  "empiriolabs",
+  "kilo",
+  "llmgateway",
+  "merge-gateway",
+  "nano-gpt",
+  "openrouter",
+  "venice",
+]);
+
+export interface CatalogChange {
+  status: "created" | "updated" | "deleted";
+  path: string;
+}
+
+export interface AutoMergeDecision {
+  safe: boolean;
+  created: number;
+  updated: number;
+  deleted: number;
+  reasons: string[];
+}
+
+function isModel(path: string) {
+  return path.endsWith(".toml") && (path.startsWith("models/") || path.includes("/models/"));
+}
+
+function isProviderModel(path: string) {
+  return path.endsWith(".toml") && path.startsWith("providers/") && path.includes("/models/");
+}
+
+export async function classifyAutoMerge(
+  changes: CatalogChange[],
+  load = (path: string) => readFile(path, "utf8"),
+  loadPrevious = load,
+): Promise<AutoMergeDecision> {
+  const models = changes.filter((change) => isModel(change.path));
+  const created = models.filter((change) => change.status === "created").length;
+  const updated = models.filter((change) => change.status === "updated").length;
+  const deleted = models.filter((change) => change.status === "deleted").length;
+  const reasons: string[] = [];
+
+  if (created > MAX_CREATED_MODELS) reasons.push(`${created} models created (limit ${MAX_CREATED_MODELS})`);
+  if (deleted > MAX_DELETED_MODELS) reasons.push(`${deleted} models deleted (limit ${MAX_DELETED_MODELS})`);
+  if (created + deleted > MAX_MODEL_CHURN) {
+    reasons.push(`${created + deleted} models created or deleted (limit ${MAX_MODEL_CHURN})`);
+  }
+
+  const reasoningMetadata = async (path: string, loader: typeof load) => {
+    const model = Bun.TOML.parse(await loader(path)) as Record<string, unknown>;
+    let reasoning = model.reasoning;
+    if (reasoning === undefined && typeof model.base_model === "string") {
+      const base = Bun.TOML.parse(await loader(`models/${model.base_model}.toml`)) as Record<string, unknown>;
+      reasoning = base.reasoning;
+    }
+
+    return {
+      reasoning,
+      reasoning_options: model.reasoning_options,
+      interleaved: model.interleaved,
+      base_model: model.base_model,
+    };
+  };
+
+  for (const change of models) {
+    if (change.status === "deleted" || !isProviderModel(change.path)) continue;
+
+    const current = await reasoningMetadata(change.path, load);
+    const previous = change.status === "created" ? undefined : await reasoningMetadata(change.path, loadPrevious);
+    const reasoningChanged = !current || !previous || !isDeepStrictEqual(current, previous);
+    if (!reasoningChanged) continue;
+
+    const reasoning = current?.reasoning === true || previous?.reasoning === true;
+
+    if (reasoning) {
+      if (current?.reasoning === true && current.reasoning_options === undefined) {
+        reasons.push(`${change.path} is a reasoning model without explicit reasoning_options`);
+      } else if (!REVIEWED_REASONING_PROVIDERS.has(change.path.split("/")[1]!)) {
+        reasons.push(`${change.path} is a reasoning model that requires manual review`);
+      }
+    }
+  }
+
+  return { safe: reasons.length === 0, created, updated, deleted, reasons };
+}
+
+export function parseNameStatus(output: string): CatalogChange[] {
+  return output.trim().split("\n").filter(Boolean).flatMap((line) => {
+    const [code, ...paths] = line.split("\t");
+    const path = paths.at(-1);
+    if (!code || !path) throw new Error(`Invalid git diff entry: ${line}`);
+    if (code.startsWith("R")) {
+      if (paths.length !== 2) throw new Error(`Invalid git rename entry: ${line}`);
+      return [
+        { status: "deleted", path: paths[0]! },
+        { status: "created", path: paths[1]! },
+      ];
+    }
+    return {
+      status: code.startsWith("A") ? "created" : code.startsWith("D") ? "deleted" : "updated",
+      path,
+    };
+  });
+}

packages/core/src/sync/index.ts

@@ -10,21 +10,26 @@ import { anthropic } from "./providers/anthropic.js";
 import { baseten } from "./providers/baseten.js";
 import { chutes } from "./providers/chutes.js";
 import { cloudflareWorkersAi } from "./providers/cloudflare-workers-ai.js";
+import { cortecs } from "./providers/cortecs.js";
 import { crossmodel } from "./providers/crossmodel.js";
 import { deepinfra } from "./providers/deepinfra.js";
 import { digitalocean } from "./providers/digitalocean.js";
+import { edenai } from "./providers/edenai.js";
 import { empiriolabs } from "./providers/empiriolabs.js";
 import { google } from "./providers/google.js";
 import { hyper } from "./providers/hyper.js";
 import { huggingface } from "./providers/huggingface.js";
+import { inceptron } from "./providers/inceptron.js";
 import { kilo } from "./providers/kilo.js";
 import { llmgateway } from "./providers/llmgateway.js";
 import { mergeGateway } from "./providers/merge-gateway.js";
 import { nanoGpt } from "./providers/nano-gpt.js";
 import { openai } from "./providers/openai.js";
+import { ofox } from "./providers/ofox.js";
 import { openrouter } from "./providers/openrouter.js";
 import { ovhcloud } from "./providers/ovhcloud.js";
 import { pioneer } from "./providers/pioneer.js";
+import { requesty } from "./providers/requesty.js";
 import { tinfoil } from "./providers/tinfoil.js";
 import { vercel } from "./providers/vercel.js";
 import { venice } from "./providers/venice.js";
@@ -114,21 +119,26 @@ export const providers: {
   baseten: SyncProvider<any>;
   chutes: SyncProvider<any>;
   "cloudflare-workers-ai": SyncProvider<any>;
+  cortecs: SyncProvider<any>;
   crossmodel: SyncProvider<any>;
   deepinfra: SyncProvider<any>;
   digitalocean: SyncProvider<any>;
+  edenai: SyncProvider<any>;
   empiriolabs: SyncProvider<any>;
   google: SyncProvider<any>;
   hyper: SyncProvider<any>;
   huggingface: SyncProvider<any>;
+  inceptron: SyncProvider<any>;
   kilo: SyncProvider<any>;
   llmgateway: SyncProvider<any>;
   "merge-gateway": SyncProvider<any>;
   "nano-gpt": SyncProvider<any>;
+  ofox: SyncProvider<any>;
   openai: SyncProvider<any>;
   openrouter: SyncProvider<any>;
   ovhcloud: SyncProvider<any>;
   pioneer: SyncProvider<any>;
+  requesty: SyncProvider<any>;
   tinfoil: SyncProvider<any>;
   vercel: SyncProvider<any>;
   venice: SyncProvider<any>;
@@ -140,21 +150,26 @@ export const providers: {
   baseten,
   chutes,
   "cloudflare-workers-ai": cloudflareWorkersAi,
+  cortecs,
   crossmodel,
   deepinfra,
   digitalocean,
+  edenai,
   empiriolabs,
   google,
   hyper,
   huggingface,
+  inceptron,
   kilo,
   llmgateway,
   "merge-gateway": mergeGateway,
   "nano-gpt": nanoGpt,
+  ofox,
   openai,
   openrouter,
   ovhcloud,
   pioneer,
+  requesty,
   tinfoil,
   vercel,
   venice,
@@ -165,17 +180,21 @@ export const providers: {
 export const groups = {
   aggregators: [
     "crossmodel",
+    "edenai",
     "empiriolabs",
     "huggingface",
+    "inceptron",
     "kilo",
     "llmgateway",
     "merge-gateway",
     "nano-gpt",
+    "ofox",
+    "requesty",
     "openrouter",
     "vercel",
   ],
   cloudflare: ["cloudflare-workers-ai"],
-  direct: ["ambient", "anthropic", "baseten", "chutes", "deepinfra", "digitalocean", "google", "hyper", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
+  direct: ["ambient", "anthropic", "baseten", "chutes", "cortecs", "deepinfra", "digitalocean", "google", "hyper", "openai", "ovhcloud", "pioneer", "tinfoil", "venice", "wandb", "xai"],
 } as const;
 
 type ProviderID = keyof typeof providers;

packages/core/src/sync/providers/chutes.ts

@@ -134,9 +134,11 @@ export function buildChutesModel(
     last_updated: existing?.last_updated ?? today,
     attachment,
     reasoning,
-    // Chutes' /v1/models advertises `reasoning` as a capability but exposes no parameter
-    // to toggle or set its effort, so there is no provider evidence for a reasoning option.
-    reasoning_options: [],
+    // Chutes' /v1/models advertises `reasoning` as a capability but lists no sampling
+    // parameter for it, so the endpoint alone cannot describe the control. The real
+    // control is `chat_template_kwargs`, which is hand-authored per model. Leaving this
+    // field unset lets preserveReasoningOptions keep those authored options and default
+    // only new, unannotated reasoners to an empty list.
     temperature,
     tool_call: toolCall,
     structured_output: structuredOutput ? true : undefined,

packages/core/src/sync/providers/cortecs.ts

@@ -0,0 +1,169 @@
+import { z } from "zod";
+
+import { describeModel } from "../../describe.js";
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel, resolveModelMetadataBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = "https://api.cortecs.ai/v1/models";
+const CANONICAL_BASE_MODEL_EXCEPTIONS = {
+  "claude-sonnet-4": "anthropic/claude-sonnet-4-0",
+} as const;
+// Cortecs publishes its default catalog prices in EUR per million tokens.
+// Exchange rate used by the existing Cortecs entries, as of 2026-07-30.
+const EUR_TO_USD = 1.114;
+
+const CortecsModality = z.enum(["text", "audio", "image", "video", "pdf"]);
+
+export const CortecsModel = z.object({
+  id: z.string().min(1),
+  created: z.number().int().nonnegative(),
+  description: z.string().optional(),
+  pricing: z.object({
+    currency: z.literal("EUR"),
+    input_token: z.number().nonnegative(),
+    output_token: z.number().nonnegative(),
+    cache_read_cost: z.number().nonnegative().optional(),
+    cache_write_cost: z.number().nonnegative().optional(),
+  }).passthrough(),
+  context_size: z.number().int().positive(),
+  input_modalities: z.array(CortecsModality).default(["text"]),
+  output_modalities: z.array(CortecsModality).default(["text"]),
+  supported_features: z.array(z.string()).default([]),
+}).passthrough();
+
+export const CortecsResponse = z.object({
+  object: z.literal("list"),
+  data: z.array(CortecsModel),
+}).passthrough();
+
+export type CortecsModel = z.infer<typeof CortecsModel>;
+
+export const cortecs = {
+  id: "cortecs",
+  name: "Cortecs",
+  modelsDir: "providers/cortecs/models",
+  deleteMissing: true,
+  async fetchModels() {
+    const response = await fetch(API_ENDPOINT);
+    if (!response.ok) {
+      throw new Error(`Cortecs models request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    return CortecsResponse.parse(raw).data;
+  },
+  translateModel(model, context) {
+    return {
+      id: model.id,
+      model: buildCortecsModel(model, context.existing(model.id), context.authored(model.id)),
+    };
+  },
+} satisfies SyncProvider<CortecsModel>;
+
+function dateFromTimestamp(timestamp: number) {
+  return new Date(timestamp * 1_000).toISOString().slice(0, 10);
+}
+
+function usd(value: number | undefined) {
+  if (value === undefined) return undefined;
+  return Math.round(value * EUR_TO_USD * 1_000) / 1_000;
+}
+
+export function buildCortecsModel(
+  model: CortecsModel,
+  existing: ExistingModel | undefined,
+  authored: ExistingModel | undefined,
+): SyncedModel {
+  const features = new Set(model.supported_features);
+  const input = model.input_modalities;
+  const output = model.output_modalities;
+  const canonical = existing?.base_model ?? resolveCortecsBaseModel(model.id);
+  const sourceReasoning = features.has("reasoning");
+  const reasoning = canonical === undefined ? sourceReasoning : existing?.reasoning ?? sourceReasoning;
+  const reasoningOptions = canonical === undefined
+    ? (sourceReasoning ? existing?.reasoning_options ?? [] : undefined)
+    : (existing?.reasoning === true ? existing.reasoning_options : undefined);
+  const limit = {
+    context: model.context_size,
+    input: existing?.limit?.input,
+    output: authored?.limit?.output ?? model.context_size,
+  };
+  const cost = {
+    input: usd(model.pricing.input_token),
+    output: usd(model.pricing.output_token),
+    cache_read: usd(model.pricing.cache_read_cost) ?? existing?.cost?.cache_read,
+    cache_write: usd(model.pricing.cache_write_cost) ?? existing?.cost?.cache_write,
+    reasoning: existing?.cost?.reasoning,
+    tiers: existing?.cost?.tiers,
+  };
+  if (canonical !== undefined) {
+    return factorBaseModel(canonical, {
+      description: existing?.description,
+      attachment: input.some((value) => value !== "text"),
+      reasoning,
+      reasoning_options: reasoningOptions,
+      temperature: existing?.temperature,
+      tool_call: features.has("tools"),
+      structured_output: features.has("json_mode"),
+      status: existing?.status,
+      interleaved: existing?.interleaved,
+      limit,
+      modalities: { input, output },
+      cost,
+    }, limit, existing?.base_model_omit);
+  }
+
+  const family = existing?.family;
+  return {
+    name: existing?.name ?? model.id,
+    description: existing?.description ?? model.description ?? describeModel({
+      id: model.id,
+      name: model.id,
+      family,
+      reasoning,
+      tool_call: features.has("tools"),
+      structured_output: features.has("json_mode"),
+      open_weights: existing?.open_weights ?? false,
+      limit,
+      modalities: { input, output },
+    }),
+    family,
+    release_date: existing?.release_date ?? dateFromTimestamp(model.created),
+    last_updated: existing?.last_updated ?? dateFromTimestamp(model.created),
+    attachment: input.some((value) => value !== "text"),
+    reasoning,
+    reasoning_options: reasoningOptions,
+    temperature: existing?.temperature ?? false,
+    tool_call: features.has("tools"),
+    structured_output: features.has("json_mode"),
+    knowledge: existing?.knowledge,
+    open_weights: existing?.open_weights ?? false,
+    status: existing?.status,
+    interleaved: existing?.interleaved,
+    cost,
+    limit,
+    modalities: { input, output },
+  } satisfies SyncedFullModel;
+}
+
+function resolveCortecsBaseModel(modelID: string) {
+  const exception = CANONICAL_BASE_MODEL_EXCEPTIONS[
+    modelID as keyof typeof CANONICAL_BASE_MODEL_EXCEPTIONS
+  ];
+  if (exception !== undefined) return resolveModelMetadataBaseModel(exception);
+
+  const trailingFamily = /^claude-(\d+)-(\d+)-(opus|sonnet|haiku)$/.exec(modelID);
+  if (trailingFamily !== null) {
+    const [, major, minor, family] = trailingFamily;
+    return resolveModelMetadataBaseModel(`anthropic/claude-${family}-${major}-${minor}`);
+  }
+
+  const compactFamily = /^claude-(opus|sonnet|haiku)(\d+)-(\d+)$/.exec(modelID);
+  if (compactFamily !== null) {
+    const [, family, major, minor] = compactFamily;
+    return resolveModelMetadataBaseModel(`anthropic/claude-${family}-${major}-${minor}`);
+  }
+
+  return resolveModelMetadataBaseModel(modelID);
+}

packages/core/src/sync/providers/crossmodel.ts

@@ -24,11 +24,13 @@ const API_ENDPOINT = process.env.CROSSMODEL_MODELS_URL ?? "https://www.crossmode
 
 const ReasoningCapability = z
   .object({
-    toggle: z.boolean().optional(),
-    effort: z.array(z.string()).optional(),
+    supported: z.boolean().optional(),
+    toggle: z.boolean().nullish().transform((value) => value ?? undefined),
+    effort: z.array(z.string()).nullish().transform((value) => value ?? undefined),
     budget_tokens: z
       .object({ min: z.number().optional(), max: z.number().optional() })
-      .optional(),
+      .nullish()
+      .transform((value) => value ?? undefined),
   })
   .passthrough();
 
@@ -55,7 +57,10 @@ export const CrossModelModel = z
       .object({ input: z.array(z.string()), output: z.array(z.string()) })
       .optional(),
     capabilities: z
-      .object({ reasoning: ReasoningCapability.optional() })
+      .object({
+        json: z.boolean().optional(),
+        reasoning: ReasoningCapability.optional(),
+      })
       .passthrough()
       .nullable()
       .optional(),
@@ -170,7 +175,7 @@ function isReasoningEffort(value: string): value is ReasoningEffort {
 //   otherwise         -> toggle / effort / budget_tokens entries
 function reasoningOptions(model: CrossModelModel): SyncedModel["reasoning_options"] {
   const reasoning = model.capabilities?.reasoning;
-  if (reasoning === undefined) return undefined;
+  if (reasoning === undefined || reasoning.supported === false) return undefined;
   const options: NonNullable<SyncedModel["reasoning_options"]> = [];
   if (reasoning.toggle === true) options.push({ type: "toggle" });
   if (reasoning.effort !== undefined) {
@@ -186,7 +191,7 @@ function reasoningOptions(model: CrossModelModel): SyncedModel["reasoning_option
   return options;
 }
 
-function buildCrossModel(
+export function buildCrossModel(
   model: CrossModelModel,
   existing: ExistingModel | undefined,
 ): SyncedModel | undefined {
@@ -247,7 +252,7 @@ function buildCrossModel(
       reasoning: existing?.reasoning,
       temperature: existing?.temperature,
       tool_call: existing?.tool_call,
-      structured_output: existing?.structured_output,
+      structured_output: model.capabilities?.json ?? existing?.structured_output,
       knowledge: existing?.knowledge,
       modalities: modality,
       reasoning_options,

packages/core/src/sync/providers/deepinfra.ts

@@ -372,6 +372,7 @@ export function buildDeepInfraModel(
 // catalog's canonical metadata namespace so new models can inherit via
 // `base_model` whenever a `models/` entry already exists.
 const DEEPINFRA_PREFIXES: Record<string, string> = {
+  ByteDance: "bytedance-seed",
   "deepseek-ai": "deepseek",
   "meta-llama": "meta",
   google: "google",

packages/core/src/sync/providers/digitalocean.ts

@@ -168,9 +168,11 @@ export const digitalocean = {
         || outputLimit <= 0
       )
     ) return undefined;
-    const baseModel = existing === undefined
-      ? resolveDigitalOceanBaseModel(model.id)
-      : existing.base_model;
+    // Only auto-resolve base_model for newly created files. Existing full
+    // definitions stay hand-authored unless they already declare base_model.
+    const baseModel = existing !== undefined
+      ? existing.base_model
+      : resolveDigitalOceanBaseModel(model.id);
     return {
       id: model.id,
       model: buildDigitalOceanModel(model, existing, baseModel),
@@ -341,6 +343,13 @@ function normalizeModalities(values: string[], fallback: Modality[]): Modality[]
   return [...new Set(normalized.length > 0 ? normalized : fallback)];
 }
 
+function normalizeEffortToken(value: string): string {
+  const normalized = value.trim().toLowerCase().replaceAll("_", "-");
+  if (normalized === "x-high" || normalized === "xhigh") return "xhigh";
+  if (normalized === "null") return "null";
+  return normalized;
+}
+
 function number(value: string | number | undefined) {
   if (value === undefined) return undefined;
   const parsed = typeof value === "number" ? value : Number.parseInt(value, 10);
@@ -360,12 +369,23 @@ function reasoningOptionsFor(
   model: DigitalOceanSourceModel,
   existing: ExistingModel | undefined,
 ): ExistingModel["reasoning_options"] {
-  if (model.reasoning_efforts === undefined) return existing?.reasoning_options;
-  const values = model.reasoning_efforts
-    .map((value) => value === "null" ? null : value)
-    .filter(isReasoningEffort);
+  if (model.reasoning_efforts === undefined || model.reasoning_efforts.length === 0) {
+    return existing?.reasoning_options;
+  }
+  const remoteValues = reasoningEfforts(model);
   const preserved = existing?.reasoning_options?.filter((option) => option.type !== "effort") ?? [];
-  return values.length > 0 ? [...preserved, { type: "effort", values }] : preserved;
+  return remoteValues.length > 0
+    ? [...preserved, { type: "effort", values: remoteValues }]
+    : existing?.reasoning_options;
+}
+
+function reasoningEfforts(model: DigitalOceanSourceModel) {
+  return (model.reasoning_efforts ?? [])
+    .map((value) => {
+      const normalized = normalizeEffortToken(value);
+      return normalized === "null" ? null : normalized;
+    })
+    .filter(isReasoningEffort);
 }
 
 function isReasoningEffort(value: string | null): value is ReasoningEffort {
@@ -384,9 +404,10 @@ function status(
   lifecycleStatus: string,
   existing: ExistingModel["status"],
 ): ExistingModel["status"] {
-  const lifecycle = lifecycleStatus.toLowerCase().replaceAll("_", "-");
+  const lifecycle = lifecycleStatus.trim().toLowerCase().replaceAll("_", "-");
+  if (lifecycle.length === 0) return existing;
   if (lifecycle === "deprecated" || lifecycle === "end-of-life") return "deprecated";
-  if (lifecycle === "public-preview") return "beta";
+  if (lifecycle === "public-preview" || lifecycle === "preview") return "beta";
   return existing === "deprecated" || existing === "beta" ? undefined : existing;
 }
 
@@ -430,16 +451,14 @@ function cost(model: DigitalOceanSourceModel, existing: ExistingModel | undefine
 export function buildDigitalOceanModel(
   model: DigitalOceanSourceModel,
   existing: ExistingModel | undefined,
-  baseModel = existing === undefined ? resolveDigitalOceanBaseModel(model.id) : existing.base_model,
+  baseModel = existing !== undefined
+    ? existing.base_model
+    : resolveDigitalOceanBaseModel(model.id),
 ): SyncedModel {
-  const input = normalizeModalities(
-    model.modalities?.input ?? [],
-    existing?.modalities?.input ?? ["text"],
-  );
-  const output = normalizeModalities(
-    model.modalities?.output ?? [],
-    existing?.modalities?.output ?? ["text"],
-  );
+  const remoteInput = normalizeModalities(model.modalities?.input ?? [], []);
+  const remoteOutput = normalizeModalities(model.modalities?.output ?? [], []);
+  const input = remoteInput.length > 0 ? remoteInput : existing?.modalities?.input ?? ["text"];
+  const output = remoteOutput.length > 0 ? remoteOutput : existing?.modalities?.output ?? ["text"];
   const context = number(model.context_window) ?? existing?.limit?.context ?? 0;
   const maxTokens = number(model.max_output_tokens ?? undefined);
   const limit = {
@@ -448,13 +467,19 @@ export function buildDigitalOceanModel(
     output: maxTokens ?? existing?.limit?.output ?? 0,
   };
   const textOutput = output.includes("text") && !output.includes("image") && !output.includes("video");
-  const remoteReasoning = textOutput
-    && ((model.thinking ?? false) || (model.reasoning_efforts?.length ?? 0) > 0);
-  const providerReasoning = remoteReasoning ? true : existing?.reasoning;
+  const remoteEfforts = reasoningEfforts(model);
+  const providerReasoning = !textOutput
+    ? existing?.reasoning
+    : model.thinking === true || remoteEfforts.length > 0
+    ? true
+    : model.thinking === false
+    ? false
+    : existing?.reasoning;
   const reasoning = providerReasoning ?? false;
-  const reasoningOptions = reasoning ? reasoningOptionsFor(model, existing) : undefined;
+  const reasoningOptions = reasoning === true ? reasoningOptionsFor(model, existing) : undefined;
   const modelStatus = status(model.lifecycle_status, existing?.status);
   const releaseDate = existing?.release_date ?? model.created_at?.slice(0, 10) ?? new Date().toISOString().slice(0, 10);
+  const attachment = input.some((value) => value !== "text");
   const values: Partial<SyncedFullModel> = {
     name: model.name,
     description: existing?.description ?? describeModel({
@@ -471,7 +496,7 @@ export function buildDigitalOceanModel(
     family: existing?.family ?? inferFamily(model.id, model.name),
     release_date: releaseDate,
     last_updated: existing?.last_updated ?? releaseDate,
-    attachment: existing?.attachment ?? input.some((value) => value !== "text"),
+    attachment,
     reasoning,
     reasoning_options: reasoningOptions,
     temperature: existing?.temperature ?? true,
@@ -492,7 +517,8 @@ export function buildDigitalOceanModel(
     return factorBaseModel(baseModel, {
       name: model.name,
       description: existing?.description,
-      attachment: input.some((value) => value !== "text"),
+      attachment,
+      modalities: { input, output },
       reasoning: providerReasoning,
       reasoning_options: reasoningOptions,
       temperature: existing?.temperature,
@@ -502,7 +528,6 @@ export function buildDigitalOceanModel(
       interleaved: existing?.interleaved,
       cost: cost(model, existing),
       limit,
-      modalities: { input, output },
       provider: existing?.provider,
       experimental: existing?.experimental,
     }, limit, existing?.base_model_omit);
@@ -537,15 +562,30 @@ export function resolveDigitalOceanBaseModel(id: string) {
   if (id.startsWith("glm-")) candidates.push(`zai/${id}`);
   if (id.startsWith("kimi-")) candidates.push(`moonshotai/${id}`);
   if (id.startsWith("minimax-")) candidates.push(`minimax/${id}`);
+  if (id.startsWith("mimo-")) {
+    const normalized = id.replace(/^mimo-v(\d+)-(\d+)/, "mimo-v$1.$2");
+    candidates.push(`xiaomi/${id}`);
+    candidates.push(`xiaomi/${normalized}`);
+  }
   if (id.startsWith("nvidia-")) candidates.push(`nvidia/${id.slice("nvidia-".length)}`);
   if (id.startsWith("alibaba-")) candidates.push(`qwen/${id.slice("alibaba-".length)}`);
   if (id.startsWith("qwen")) candidates.push(`qwen/${id}`);
   if (id.startsWith("llama")) candidates.push(`meta/${id}`);
   if (id.startsWith("mistral") || id.startsWith("ministral")) candidates.push(`mistralai/${id}`);
+  if (id.startsWith("gemma")) candidates.push(`google/${id}`);
 
-  const anthropic = id.match(/^anthropic-claude-(\d+(?:\.\d+)?)-(opus|sonnet|haiku)$/);
-  if (anthropic !== null) {
-    candidates.push(`anthropic/claude-${anthropic[2]}-${anthropic[1]}`);
+  // anthropic-claude-5-sonnet → anthropic/claude-sonnet-5
+  const anthropicSwapped = id.match(/^anthropic-claude-(\d+(?:\.\d+)?)-(opus|sonnet|haiku)$/);
+  if (anthropicSwapped !== null) {
+    candidates.push(`anthropic/claude-${anthropicSwapped[2]}-${anthropicSwapped[1]}`);
+  }
+  // anthropic-claude-opus-5 → anthropic/claude-opus-5
+  // also normalize dotted versions: anthropic-claude-opus-4.6 → anthropic/claude-opus-4-6
+  const anthropicFamily = id.match(/^anthropic-claude-(opus|sonnet|haiku)-(\d+(?:\.\d+)?)$/);
+  if (anthropicFamily !== null) {
+    const version = anthropicFamily[2].replaceAll(".", "-");
+    candidates.push(`anthropic/claude-${anthropicFamily[1]}-${anthropicFamily[2]}`);
+    candidates.push(`anthropic/claude-${anthropicFamily[1]}-${version}`);
   }
   if (id.startsWith("anthropic-")) candidates.push(`anthropic/${id.slice("anthropic-".length)}`);
 

packages/core/src/sync/providers/edenai.ts

@@ -0,0 +1,506 @@
+import { existsSync, readdirSync, readFileSync } from "node:fs";
+import path from "node:path";
+
+import { z } from "zod";
+
+import type { SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import {
+  factorBaseModel,
+  modelMetadata,
+  resolveModelMetadataBaseModel,
+} from "./openrouter.js";
+
+// ========================================
+// Constants
+// ========================================
+
+const API_ENDPOINT = "https://api.edenai.run/v3/models";
+const MODELS_DIR = path.join(
+  import.meta.dirname,
+  "..",
+  "..",
+  "..",
+  "..",
+  "..",
+  "models",
+);
+const PROVIDERS_DIR = path.join(MODELS_DIR, "..", "providers");
+const TOKENS_PER_MILLION = 1_000_000;
+const PRICE_DECIMALS = 1_000_000;
+
+// Values `reasoning_effort` accepts on POST /v3/chat/completions. Which of them
+// a given model exposes comes from its lab entry, not from this list.
+const ACCEPTED_EFFORTS = new Set([
+  "none",
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max",
+]);
+
+const REGION_SUFFIX = /@[a-z0-9-]+$/i;
+const DOTTED_VENDOR = /^[a-z0-9-]+\./;
+const VERSION_TAIL = /-v\d+:\d+$/;
+const DATE_TAIL = /-\d{8}$/;
+const DATABRICKS_PREFIX = "databricks-";
+const TIER_KEY = /^input_cost_per_token_above_(\d+)k_tokens$/;
+
+const MODALITY_BY_EDENAI: Record<
+  string,
+  SyncedFullModel["modalities"]["input"][number]
+> = {
+  text: "text",
+  image: "image",
+  audio: "audio",
+  video: "video",
+  file: "pdf",
+};
+
+// Upstreams that are the lab's own API for models under that namespace.
+const LAB_UPSTREAMS: Record<string, readonly string[]> = {
+  alibaba: ["qwen"],
+  amazon: ["amazon"],
+  anthropic: ["anthropic"],
+  cohere: ["cohere"],
+  deepseek: ["deepseek"],
+  google: ["google", "vertex"],
+  microsoft: ["microsoft"],
+  minimax: ["minimax"],
+  mistral: ["mistral"],
+  moonshotai: ["moonshot"],
+  openai: ["openai"],
+  perplexity: ["perplexityai"],
+  xai: ["xai"],
+  zhipuai: ["zai"],
+};
+
+type ReasoningOption = NonNullable<
+  SyncedFullModel["reasoning_options"]
+>[number];
+
+const canonicalNameByID = new Map<string, string>();
+
+let firstPartyBaseModels: ReadonlySet<string> = new Set();
+
+// ========================================
+// Schemas
+// ========================================
+
+const EdenAIPricing = z
+  .object({
+    input_cost_per_token: z.number().nullish(),
+    output_cost_per_token: z.number().nullish(),
+    output_cost_per_reasoning_token: z.number().nullish(),
+    cache_read_input_token_cost: z.number().nullish(),
+    cache_creation_input_token_cost: z.number().nullish(),
+    input_cost_per_audio_token: z.number().nullish(),
+  })
+  .passthrough();
+
+const EdenAICapabilities = z
+  .object({
+    input_modalities: z.array(z.string()).nullish(),
+    output_modalities: z.array(z.string()).nullish(),
+    supports_function_calling: z.boolean().optional(),
+    supports_response_schema: z.boolean().optional(),
+  })
+  .passthrough();
+
+export const EdenAIModel = z
+  .object({
+    id: z.string().min(1),
+    owned_by: z.string().min(1),
+    model_name: z.string().min(1),
+    context_length: z.number().nullish(),
+    capabilities: EdenAICapabilities,
+    pricing: EdenAIPricing.nullish(),
+    list_pricing: EdenAIPricing.nullish(),
+    alias_of: z.string().nullish(),
+  })
+  .passthrough();
+
+export const EdenAIResponse = z
+  .object({
+    object: z.literal("list"),
+    data: z.array(EdenAIModel),
+  })
+  .passthrough();
+
+export type EdenAIModel = z.infer<typeof EdenAIModel>;
+
+// ========================================
+// Base model resolution
+// ========================================
+
+function baseModelExists(modelID: string) {
+  return existsSync(path.join(MODELS_DIR, `${modelID}.toml`));
+}
+
+// Ids are `<upstream>/<native id>`, so each upstream keeps its own convention.
+function baseModelCandidates(modelName: string) {
+  const candidates = [modelName];
+
+  if (DOTTED_VENDOR.test(modelName)) {
+    const dotted = modelName.replace(".", "/");
+    candidates.push(
+      dotted,
+      dotted.replace(VERSION_TAIL, "").replace(DATE_TAIL, ""),
+    );
+  }
+
+  const last = modelName.split("/").at(-1) ?? modelName;
+  candidates.push(last);
+  if (last.startsWith(DATABRICKS_PREFIX)) {
+    candidates.push(last.slice(DATABRICKS_PREFIX.length));
+  }
+  candidates.push(last.replace(VERSION_TAIL, "").replace(DATE_TAIL, ""));
+
+  return [...new Set(candidates)].filter((candidate) => candidate.length > 0);
+}
+
+export function resolveEdenAIBaseModel(model: EdenAIModel) {
+  const names = [model.model_name.replace(REGION_SUFFIX, "")];
+  if (model.alias_of != null) {
+    const target = model.alias_of.split("/").slice(1).join("/");
+    if (target.length > 0) names.push(target);
+  }
+
+  for (const name of names) {
+    for (const candidate of baseModelCandidates(name)) {
+      const resolved = resolveModelMetadataBaseModel(candidate);
+      if (resolved !== undefined && baseModelExists(resolved)) return resolved;
+    }
+  }
+  return undefined;
+}
+
+function isFirstPartyRoute(model: EdenAIModel, baseModel: string) {
+  const lab = baseModel.split("/")[0] ?? "";
+  return (LAB_UPSTREAMS[lab] ?? []).includes(model.owned_by);
+}
+
+export function collectFirstPartyBaseModels(models: readonly EdenAIModel[]) {
+  const bases = new Set<string>();
+  for (const model of models) {
+    const baseModel = resolveEdenAIBaseModel(model);
+    if (baseModel !== undefined && isFirstPartyRoute(model, baseModel)) {
+      bases.add(baseModel);
+    }
+  }
+  return bases;
+}
+
+function canonicalModelName(baseModel: string) {
+  let cached = canonicalNameByID.get(baseModel);
+  if (cached === undefined) {
+    try {
+      const toml = Bun.TOML.parse(
+        readFileSync(path.join(MODELS_DIR, `${baseModel}.toml`), "utf8"),
+      ) as { name?: unknown };
+      cached = typeof toml.name === "string" ? toml.name : "";
+    } catch {
+      cached = "";
+    }
+    canonicalNameByID.set(baseModel, cached);
+  }
+  return cached === "" ? undefined : cached;
+}
+
+function hasOutputLimit(baseModel: string) {
+  const limit = modelMetadata(baseModel).limit;
+  return (
+    typeof limit === "object" &&
+    limit !== null &&
+    typeof (limit as { output?: unknown }).output === "number"
+  );
+}
+
+function regionVariantName(model: EdenAIModel, baseModel: string) {
+  const region = REGION_SUFFIX.exec(model.id)?.[0].slice(1);
+  if (region === undefined) return undefined;
+
+  const canonical = canonicalModelName(baseModel);
+  if (canonical === undefined) return undefined;
+  return `${canonical} (${region.toUpperCase()})`;
+}
+
+// ========================================
+// Reasoning options
+// ========================================
+
+// Eden AI's only reasoning control is `reasoning_effort`, so a model is
+// published with the effort list its lab entry (or an established relay peer)
+// already documents. Peers exposing only `toggle` / `budget_tokens` have no
+// equivalent here, and those models are skipped rather than given a guess.
+function effortValues(options: unknown): string[] | "always-on" | undefined {
+  if (!Array.isArray(options)) return undefined;
+  if (options.length === 0) return "always-on";
+
+  let toggled = false;
+  let accepted: string[] | undefined;
+
+  for (const option of options) {
+    if (typeof option !== "object" || option === null) continue;
+    const type = (option as { type?: unknown }).type;
+    if (type === "toggle") toggled = true;
+    if (type !== "effort") continue;
+
+    const values = (option as { values?: unknown }).values;
+    if (!Array.isArray(values)) continue;
+    const filtered = values.filter(
+      (value): value is string =>
+        typeof value === "string" && ACCEPTED_EFFORTS.has(value),
+    );
+    if (filtered.length > 0) accepted = filtered;
+  }
+
+  if (accepted === undefined) return undefined;
+  // Eden AI switches reasoning off with `reasoning_effort=none`, so a lab-side
+  // toggle becomes `none` in the effort list instead of a separate option.
+  return toggled && !accepted.includes("none")
+    ? ["none", ...accepted]
+    : accepted;
+}
+
+function parseToml(filePath: string) {
+  try {
+    return Bun.TOML.parse(readFileSync(filePath, "utf8")) as Record<
+      string,
+      unknown
+    >;
+  } catch {
+    return undefined;
+  }
+}
+
+function tomlFilesIn(dir: string): string[] {
+  let entries;
+  try {
+    entries = readdirSync(dir, { withFileTypes: true });
+  } catch {
+    return [];
+  }
+  return entries.flatMap((entry) =>
+    entry.isDirectory()
+      ? tomlFilesIn(path.join(dir, entry.name))
+      : entry.name.endsWith(".toml")
+        ? [path.join(dir, entry.name)]
+        : [],
+  );
+}
+
+// OpenRouter is the established same-surface relay, so it is the one peer
+// consulted when a lab entry documents no effort levels.
+const PEER_PROVIDER = "openrouter";
+
+let peerEfforts: Map<string, string[] | "always-on"> | undefined;
+
+function peerEffortsByBaseModel() {
+  if (peerEfforts !== undefined) return peerEfforts;
+
+  peerEfforts = new Map();
+  for (const file of tomlFilesIn(
+    path.join(PROVIDERS_DIR, PEER_PROVIDER, "models"),
+  )) {
+    const toml = parseToml(file);
+    const base = toml?.base_model;
+    if (typeof base !== "string" || peerEfforts.has(base)) continue;
+    const values = effortValues(toml?.reasoning_options);
+    if (values !== undefined) peerEfforts.set(base, values);
+  }
+  return peerEfforts;
+}
+
+export function reasoningOptionsFor(
+  baseModel: string,
+): SyncedFullModel["reasoning_options"] | undefined {
+  const [lab, ...rest] = baseModel.split("/");
+  const firstParty = effortValues(
+    parseToml(
+      path.join(PROVIDERS_DIR, lab ?? "", "models", `${rest.join("/")}.toml`),
+    )?.reasoning_options,
+  );
+  const derived = firstParty ?? peerEffortsByBaseModel().get(baseModel);
+
+  if (derived === undefined) return undefined;
+  if (derived === "always-on") return [];
+  return [{ type: "effort", values: derived } as ReasoningOption];
+}
+
+// ========================================
+// Cost
+// ========================================
+
+function pricePerMillion(price: number) {
+  return (
+    Math.round(price * TOKENS_PER_MILLION * PRICE_DECIMALS) / PRICE_DECIMALS
+  );
+}
+
+function chargedPricePerMillion(price: unknown) {
+  return typeof price === "number" && price > 0
+    ? pricePerMillion(price)
+    : undefined;
+}
+
+function costTiers(pricing: Record<string, unknown>) {
+  const thresholds = Object.keys(pricing)
+    .map((key) => TIER_KEY.exec(key)?.[1])
+    .filter((value): value is string => value !== undefined)
+    .map(Number)
+    .sort((a, b) => a - b);
+
+  return thresholds.flatMap((threshold) => {
+    // Built explicitly so `..._above_1hr_above_200k_tokens` is never read as a
+    // context tier.
+    const suffix = `_above_${threshold}k_tokens`;
+    const input = pricing[`input_cost_per_token${suffix}`];
+    const output = pricing[`output_cost_per_token${suffix}`];
+    if (typeof input !== "number" || typeof output !== "number") return [];
+
+    return [
+      {
+        tier: { type: "context" as const, size: threshold * 1_000 },
+        input: pricePerMillion(input),
+        output: pricePerMillion(output),
+        cache_read: chargedPricePerMillion(
+          pricing[`cache_read_input_token_cost${suffix}`],
+        ),
+        cache_write: chargedPricePerMillion(
+          pricing[`cache_creation_input_token_cost${suffix}`],
+        ),
+      },
+    ];
+  });
+}
+
+function buildCost(
+  model: EdenAIModel,
+  reasoning: boolean,
+): SyncedFullModel["cost"] {
+  // `pricing` carries account-level discounts; `list_pricing` is the public rate.
+  const pricing = model.list_pricing ?? model.pricing;
+  if (pricing == null) return undefined;
+
+  const input = pricing.input_cost_per_token;
+  const output = pricing.output_cost_per_token;
+  if (input == null || output == null) return undefined;
+
+  const tiers = costTiers(pricing);
+  return {
+    input: pricePerMillion(input),
+    output: pricePerMillion(output),
+    reasoning: reasoning
+      ? chargedPricePerMillion(pricing.output_cost_per_reasoning_token)
+      : undefined,
+    cache_read: chargedPricePerMillion(pricing.cache_read_input_token_cost),
+    cache_write: chargedPricePerMillion(
+      pricing.cache_creation_input_token_cost,
+    ),
+    input_audio: chargedPricePerMillion(pricing.input_cost_per_audio_token),
+    tiers: tiers.length > 0 ? tiers : undefined,
+  };
+}
+
+// ========================================
+// Model translation
+// ========================================
+
+function mapModalities(values: readonly string[] | null | undefined) {
+  if (values == null) return undefined;
+
+  const mapped = [
+    ...new Set(
+      values
+        .map((value) => MODALITY_BY_EDENAI[value.toLowerCase()])
+        .filter(
+          (value): value is NonNullable<typeof value> => value !== undefined,
+        ),
+    ),
+  ];
+  return mapped.length > 0 ? mapped : undefined;
+}
+
+export function buildEdenAIModel(
+  model: EdenAIModel,
+  firstParty: ReadonlySet<string> = firstPartyBaseModels,
+): SyncedModel | undefined {
+  const baseModel = resolveEdenAIBaseModel(model);
+  // Eden AI relays other labs' models only, so an entry needs its lab metadata.
+  if (baseModel === undefined) return undefined;
+  // The catalog reports no output limit, so the base has to resolve one.
+  if (!hasOutputLimit(baseModel)) return undefined;
+  // Where Eden AI relays the lab's own API, that route is the entry. Models
+  // with no first-party route keep every route, since their prices differ and
+  // there is no canonical one to pick.
+  if (firstParty.has(baseModel) && !isFirstPartyRoute(model, baseModel)) {
+    return undefined;
+  }
+
+  const capabilities = model.capabilities;
+  const input = mapModalities(capabilities.input_modalities);
+  const output = mapModalities(capabilities.output_modalities);
+  const modalities =
+    input !== undefined && output !== undefined ? { input, output } : undefined;
+
+  // Whether a model reasons is a property of the model, not of the relay, so
+  // the lab entry owns it and only the effort controls are authored here.
+  const reasoning = modelMetadata(baseModel).reasoning === true;
+  const reasoningOptions = reasoning
+    ? reasoningOptionsFor(baseModel)
+    : undefined;
+  if (reasoning && reasoningOptions === undefined) return undefined;
+
+  const limit =
+    model.context_length != null && model.context_length > 0
+      ? { context: model.context_length }
+      : undefined;
+
+  return factorBaseModel(
+    baseModel,
+    {
+      name: regionVariantName(model, baseModel),
+      modalities,
+      attachment: input?.some((value) => value !== "text"),
+      reasoning_options: reasoningOptions,
+      tool_call: capabilities.supports_function_calling,
+      structured_output: capabilities.supports_response_schema,
+      cost: buildCost(model, reasoning),
+      limit,
+    },
+    limit,
+  );
+}
+
+// ========================================
+// Eden AI provider
+// ========================================
+
+export const edenai = {
+  id: "edenai",
+  name: "Eden AI",
+  modelsDir: "providers/edenai/models",
+  preserveBaseModels: false,
+  preserveDescriptions: false,
+  async fetchModels() {
+    const response = await fetch(API_ENDPOINT);
+    if (!response.ok) {
+      throw new Error(
+        `Eden AI request failed: ${response.status} ${response.statusText}`,
+      );
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    const models = EdenAIResponse.parse(raw).data;
+    firstPartyBaseModels = collectFirstPartyBaseModels(models);
+    return models;
+  },
+  translateModel(model) {
+    const built = buildEdenAIModel(model);
+    if (built === undefined) return undefined;
+    return { id: model.id, model: built };
+  },
+} satisfies SyncProvider<EdenAIModel>;

packages/core/src/sync/providers/empiriolabs.ts

@@ -1,26 +1,18 @@
 import { z } from "zod";
 
 import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
-import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
+import { factorBaseModel, resolveCanonicalBaseModel, resolveModelMetadataBaseModel } from "./openrouter.js";
 
 // EmpirioLabs exposes a public, unauthenticated OpenAI-compatible model
 // catalog, so no API key is needed or used for this sync.
 const API_ENDPOINT = "https://api.empiriolabs.ai/v1/models";
 
+// Keep this for slugs that cannot be derived from a lab filename.
+// Family prefixes, version-dot slugs, unique filenames, and dated/version
+// suffixes are resolved automatically by resolveEmpiriolabsBaseModel.
 const CANONICAL_BASE_MODELS: Record<string, string> = {
-  "fugu-ultra": "sakana/fugu-ultra",
-  "deepseek-v4-flash-0731": "deepseek/deepseek-v4-flash",
-  "gemma-4-26b-a4b": "google/gemma-4-26b-a4b-it",
-  "gemma-4-e4b": "google/gemma-4-E4B-it",
   "mistral-medium-3": "mistral/mistral-medium-2505",
   "mistral-small-4": "mistral/mistral-small-2603",
-  "muse-spark-1-1": "meta/muse-spark-1.1",
-  "qwen3-5-9b": "alibaba/qwen3.5-9b",
-  "qwen3-7-max": "alibaba/qwen3.7-max",
-  "qwen3-7-plus": "alibaba/qwen3.7-plus",
-  "step-3-5-flash": "stepfun/step-3.5-flash",
-  "step-3-5-flash-2603": "stepfun/step-3.5-flash-2603",
-  "step-3-7-flash": "stepfun/step-3.7-flash",
 };
 
 const EmpiriolabsParameter = z
@@ -209,55 +201,75 @@ function parameterOutputLimit(model: EmpiriolabsModel) {
   return parameter?.max !== undefined && parameter.max > 0 ? parameter.max : undefined;
 }
 
+function applyVersionDots(id: string) {
+  return id
+    .replace(/^(qwen\d+)-(\d+)/, "$1.$2")
+    .replace(/^(seed-\d+)-(\d+)/, "$1.$2")
+    .replace(/^(muse-[a-z]+)-(\d+)-(\d+)$/, "$1-$2.$3")
+    .replace(/^(glm-\d+)-(\d+)/, "$1.$2")
+    .replace(/^(kimi-k\d+)-(\d+)/, "$1.$2")
+    .replace(/^(minimax-m\d+)-(\d+)/, "$1.$2")
+    .replace(/^(mimo-v\d+)-(\d+)/, "$1.$2")
+    .replace(/^(deepseek-v\d+)-(\d+)/, "$1.$2")
+    .replace(/^(step-\d+)-(\d+)/, "$1.$2");
+}
+
+function stripProductSuffixes(id: string) {
+  const out: string[] = [];
+  if (/-v\d+(-\d+)?$/.test(id)) {
+    const dropPatch = id.replace(/-\d+$/, "");
+    if (dropPatch !== id) out.push(dropPatch);
+    out.push(id.replace(/-v\d+(-\d+)?$/, ""));
+  }
+  if (/-\d{4}$/.test(id)) out.push(id.replace(/-\d{4}$/, ""));
+  return out;
+}
+
+function idVariants(id: string) {
+  const variants = [id];
+  const dotted = applyVersionDots(id);
+  if (dotted !== id) variants.push(dotted);
+  for (const stripped of stripProductSuffixes(id)) {
+    if (!variants.includes(stripped)) variants.push(stripped);
+    const strippedDotted = applyVersionDots(stripped);
+    if (!variants.includes(strippedDotted)) variants.push(strippedDotted);
+  }
+  return variants;
+}
+
+function prefixesFor(id: string) {
+  if (id.startsWith("deepseek-")) return ["deepseek"];
+  if (id.startsWith("glm-")) return ["z-ai"];
+  if (id.startsWith("kimi-")) return ["moonshotai"];
+  if (id.startsWith("minimax-")) return ["minimax"];
+  if (id.startsWith("mimo-")) return ["xiaomi"];
+  if (id.startsWith("qwen")) return ["qwen"];
+  if (id.startsWith("muse-")) return ["meta"];
+  if (id.startsWith("seed-")) return ["bytedance-seed"];
+  if (id.startsWith("fugu-")) return ["sakana"];
+  if (id.startsWith("gemma-")) return ["google"];
+  if (id.startsWith("step") && !id.startsWith("stepaudio")) return ["stepfun"];
+  if (id.startsWith("mistral-")) return ["mistralai"];
+  return [];
+}
+
 export function resolveEmpiriolabsBaseModel(id: string) {
   const explicit = CANONICAL_BASE_MODELS[id];
   if (explicit !== undefined) return explicit;
-  return canonicalCandidates(id)
-    .map((candidate) => resolveCanonicalBaseModel(candidate))
-    .find((candidate) => candidate !== undefined);
-}
 
-function canonicalCandidates(id: string) {
-  const candidates: string[] = [];
-
-  if (id.startsWith("deepseek-")) {
-    candidates.push(`deepseek/${id}`);
-    candidates.push(`deepseek/${id.replace(/^deepseek-v(\d+)-(\d+)/, "deepseek-v$1.$2")}`);
+  for (const variant of idVariants(id)) {
+    for (const prefix of prefixesFor(variant)) {
+      const resolved = resolveCanonicalBaseModel(`${prefix}/${variant}`);
+      if (resolved !== undefined) return resolved;
+      if (prefix === "google" && !variant.endsWith("-it")) {
+        const instruct = resolveCanonicalBaseModel(`${prefix}/${variant}-it`);
+        if (instruct !== undefined) return instruct;
+      }
+    }
+    const unique = resolveModelMetadataBaseModel(variant);
+    if (unique !== undefined) return unique;
   }
-
-  if (id.startsWith("glm-")) {
-    const normalized = id
-      .replace(/^glm-(\d+)-(\d+)/, "glm-$1.$2")
-      .replace(/^glm-(\d+)-(\d+)v/, "glm-$1.$2v");
-    candidates.push(`z-ai/${id}`);
-    candidates.push(`z-ai/${normalized}`);
-  }
-
-  if (id.startsWith("kimi-")) {
-    const normalized = id.replace(/^(kimi-k\d+)-(\d+)/, "$1.$2");
-    candidates.push(`moonshotai/${id}`);
-    candidates.push(`moonshotai/${normalized}`);
-  }
-
-  if (id.startsWith("minimax-")) {
-    const normalized = id.replace(/^minimax-m(\d+)-(\d+)/, "minimax-m$1.$2");
-    candidates.push(`minimax/${id}`);
-    candidates.push(`minimax/${normalized}`);
-  }
-
-  if (id.startsWith("mimo-")) {
-    const normalized = id.replace(/^mimo-v(\d+)-(\d+)/, "mimo-v$1.$2");
-    candidates.push(`xiaomi/${id}`);
-    candidates.push(`xiaomi/${normalized}`);
-  }
-
-  if (id.startsWith("qwen")) {
-    const normalized = id.replace(/^(qwen\d+)-(\d+)/, "$1.$2");
-    candidates.push(`qwen/${id}`);
-    candidates.push(`qwen/${normalized}`);
-  }
-
-  return [...new Set(candidates)];
+  return undefined;
 }
 
 export function buildEmpiriolabsModel(

packages/core/src/sync/providers/hyper.ts

@@ -4,7 +4,7 @@ import { z } from "zod";
 
 import { describeModel } from "../../describe.js";
 import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
-import { factorBaseModel, resolveModelMetadataBaseModel } from "./openrouter.js";
+import { factorBaseModel, modelMetadata, resolveModelMetadataBaseModel } from "./openrouter.js";
 
 const API_ENDPOINT = "https://hyper.charm.land/v1/models";
 const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
@@ -145,7 +145,16 @@ export function buildHyperModel(
     output: model.max_output_tokens,
   };
   const modalities = hyperModalities(model.capabilities?.vision ?? false);
-  const reasoning = model.reasoning != null;
+  const resolvedBase = resolveHyperBaseModel(model.id, baseModel);
+  const advertisedReasoning = model.reasoning != null;
+  // An omitted reasoning object means Hyper advertises no controls, not that a
+  // canonical model's reasoning capability is disabled.
+  const reasoning = resolvedBase !== undefined && !advertisedReasoning
+    ? undefined
+    : advertisedReasoning;
+  const inheritedReasoning = resolvedBase === undefined
+    ? false
+    : modelMetadata(resolvedBase).reasoning === true;
   const releaseDate = existing?.release_date ?? dateFromTimestamp(model.created);
   const values: Partial<SyncedFullModel> = {
     attachment: modalities.input.some((value) => value !== "text"),
@@ -157,9 +166,12 @@ export function buildHyperModel(
     cost: buildCost(model, existing?.cost),
     limit,
   };
-  if (reasoning) values.reasoning_options = reasoningOptions(model);
+  if (advertisedReasoning) {
+    values.reasoning_options = reasoningOptions(model);
+  } else if (inheritedReasoning) {
+    values.reasoning_options = [];
+  }
 
-  const resolvedBase = resolveHyperBaseModel(model.id, baseModel);
   if (resolvedBase !== undefined) {
     return factorBaseModel(
       resolvedBase,

packages/core/src/sync/providers/inceptron.ts

@@ -0,0 +1,208 @@
+import { existsSync } from "node:fs";
+import path from "node:path";
+
+import { z } from "zod";
+
+import { ReasoningOption } from "../../schema.js";
+import type { SyncProvider, SyncedModel } from "../index.js";
+import { factorBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = process.env.INCEPTRON_MODELS_URL ?? "https://api.inceptron.io/v1/models";
+const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
+
+const ModelsDevMetadata = z
+  .object({
+    base_model: z.string().regex(/^[^./\\][^/\\]*\/[^./\\][^/\\]*$/),
+    reasoning_options: z.array(ReasoningOption).optional(),
+    interleaved: z
+      .union([
+        z.literal(true),
+        z.object({ field: z.enum(["reasoning_content", "reasoning_details"]) }).strict(),
+      ])
+      .optional(),
+    status: z.enum(["alpha", "beta", "deprecated"]).optional(),
+  })
+  .strict();
+
+export const InceptronModel = z.object({
+  id: z.string().min(1),
+  name: z.string().min(1),
+  is_ready: z.boolean().optional(),
+  context_length: z.number().int().positive(),
+  max_output_length: z.number().int().positive(),
+  input_modalities: z.array(z.string()).min(1),
+  output_modalities: z.array(z.string()).min(1),
+  supported_features: z.array(z.string()),
+  supported_sampling_parameters: z.array(z.string()),
+  pricing: z.object({
+    prompt: z.string(),
+    completion: z.string(),
+    input_cache_reads: z.string().optional(),
+    input_cache_writes: z.string().optional(),
+  }),
+  models_dev: ModelsDevMetadata.optional(),
+});
+
+export const InceptronResponse = z
+  .object({
+    object: z.literal("list"),
+    data: z.array(InceptronModel),
+  })
+  .strict();
+
+export type InceptronModel = z.infer<typeof InceptronModel>;
+export type ReadyInceptronModel = InceptronModel & {
+  models_dev: z.infer<typeof ModelsDevMetadata>;
+};
+
+export const inceptron = {
+  id: "inceptron",
+  name: "Inceptron",
+  modelsDir: "providers/inceptron/models",
+  async fetchModels() {
+    const response = await fetch(API_ENDPOINT);
+    if (!response.ok) {
+      throw new Error(`Inceptron request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels: parseInceptronModels,
+  translateModel(model) {
+    return { id: model.id, model: buildInceptronModel(model) };
+  },
+} satisfies SyncProvider<ReadyInceptronModel>;
+
+export function parseInceptronModels(raw: unknown): ReadyInceptronModel[] {
+  const models = InceptronResponse.parse(raw).data.filter((model) => model.is_ready !== false);
+  const seen = new Set<string>();
+
+  return models.map((model) => {
+    if (seen.has(model.id)) throw new Error(`Duplicate ready Inceptron model ID: ${model.id}`);
+    seen.add(model.id);
+    if (model.models_dev === undefined) {
+      throw new Error(`Ready Inceptron model ${model.id} is missing models_dev metadata`);
+    }
+    if (!baseModelExists(model.models_dev.base_model)) {
+      throw new Error(
+        `Ready Inceptron model ${model.id} refers to missing base model ${model.models_dev.base_model}`,
+      );
+    }
+    validateReasoningContract(model as ReadyInceptronModel);
+    validatePricing(model);
+    validateModalities(model);
+    return model as ReadyInceptronModel;
+  });
+}
+
+export function buildInceptronModel(model: ReadyInceptronModel): SyncedModel {
+  const features = new Set(model.supported_features);
+  const samplingParameters = new Set(model.supported_sampling_parameters);
+  const input = validateModalities(model).input;
+  const output = validateModalities(model).output;
+  const limit = {
+    context: model.context_length,
+    output: model.max_output_length,
+  };
+
+  return factorBaseModel(
+    model.models_dev.base_model,
+    {
+      name: model.name,
+      attachment: input.some((modality) => modality !== "text"),
+      reasoning: features.has("reasoning"),
+      reasoning_options: model.models_dev.reasoning_options,
+      interleaved: model.models_dev.interleaved,
+      tool_call: features.has("tools"),
+      structured_output: features.has("structured_outputs"),
+      temperature: samplingParameters.has("temperature"),
+      status: model.models_dev.status,
+      cost: {
+        input: perTokenToPerMillion(model.pricing.prompt),
+        output: perTokenToPerMillion(model.pricing.completion),
+        cache_read: optionalPrice(model.pricing.input_cache_reads),
+        cache_write: optionalPrice(model.pricing.input_cache_writes),
+      },
+      limit,
+      modalities: { input, output },
+    },
+    limit,
+  );
+}
+
+function baseModelExists(modelID: string) {
+  return existsSync(path.join(MODELS_DIR, `${modelID}.toml`));
+}
+
+function validateReasoningContract(model: ReadyInceptronModel) {
+  const supportsReasoning = model.supported_features.includes("reasoning");
+  const options = model.models_dev.reasoning_options;
+  if (supportsReasoning !== (options !== undefined)) {
+    throw new Error(
+      `Inceptron model ${model.id} must expose reasoning_options exactly when reasoning is supported`,
+    );
+  }
+  if (model.models_dev.interleaved !== undefined && !supportsReasoning) {
+    throw new Error(`Inceptron model ${model.id} exposes interleaving without reasoning`);
+  }
+
+  const optionTypes = options?.map((option) => option.type) ?? [];
+  if (new Set(optionTypes).size !== optionTypes.length) {
+    throw new Error(`Inceptron model ${model.id} has duplicate reasoning option types`);
+  }
+  const exposesEffort = optionTypes.includes("effort");
+  const advertisesEffort = model.supported_sampling_parameters.includes("reasoning_effort");
+  if (exposesEffort !== advertisesEffort) {
+    throw new Error(
+      `Inceptron model ${model.id} must advertise reasoning_effort exactly when effort options are exposed`,
+    );
+  }
+}
+
+type Modality = "text" | "audio" | "image" | "video" | "pdf";
+const MODALITIES = new Set<Modality>(["text", "audio", "image", "video", "pdf"]);
+
+function validateModalities(model: InceptronModel): { input: Modality[]; output: Modality[] } {
+  const parse = (direction: "input" | "output", values: string[]) => {
+    const unique = [...new Set(values)];
+    for (const value of unique) {
+      if (!MODALITIES.has(value as Modality)) {
+        throw new Error(`Inceptron model ${model.id} has unsupported ${direction} modality: ${value}`);
+      }
+    }
+    return unique as Modality[];
+  };
+  return {
+    input: parse("input", model.input_modalities),
+    output: parse("output", model.output_modalities),
+  };
+}
+
+function validatePricing(model: InceptronModel) {
+  perTokenToPerMillion(model.pricing.prompt);
+  perTokenToPerMillion(model.pricing.completion);
+  optionalPrice(model.pricing.input_cache_reads);
+  optionalPrice(model.pricing.input_cache_writes);
+}
+
+function optionalPrice(value: string | undefined) {
+  return value === undefined ? undefined : perTokenToPerMillion(value);
+}
+
+/** Convert a non-negative decimal USD/token string to USD/million tokens without floating-point multiplication. */
+export function perTokenToPerMillion(value: string): number {
+  const match = /^(0|[1-9]\d*)(?:\.(\d+))?$/.exec(value);
+  if (match === null) throw new Error(`Invalid Inceptron per-token price: ${value}`);
+
+  const integer = match[1] as string;
+  const fraction = match[2] ?? "";
+  const digits = `${integer}${fraction}`.replace(/^0+(?=\d)/, "");
+  const decimalPlaces = fraction.length - 6;
+  const scaled = decimalPlaces <= 0
+    ? `${digits}${"0".repeat(-decimalPlaces)}`
+    : `${digits.slice(0, -decimalPlaces) || "0"}.${digits.slice(-decimalPlaces).padStart(decimalPlaces, "0")}`;
+  const result = Number(scaled);
+  if (!Number.isFinite(result) || result < 0) {
+    throw new Error(`Invalid Inceptron per-token price: ${value}`);
+  }
+  return result;
+}

packages/core/src/sync/providers/kilo.ts

@@ -1,17 +1,10 @@
 import { z } from "zod";
-import { readFileSync, readdirSync } from "node:fs";
-import path from "node:path";
-
 import { describeModel } from "../../describe.js";
 import { inferKimiFamily, ModelFamilyValues } from "../../family.js";
 import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
 import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
 
 const API_ENDPOINT = "https://api.kilo.ai/api/gateway/models";
-const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
-const modelMetadataByID = new Map<string, Record<string, unknown>>();
-const modelMetadataFilesByProvider = new Map<string, Set<string>>();
-
 
 export const KiloModel = z.object({
   id: z.string(),
@@ -158,9 +151,9 @@ export function buildKiloModel(
   const prompt = price(model.pricing.prompt);
   const completion = price(model.pricing.completion);
   const reasoning = params.has("reasoning") || params.has("include_reasoning");
-  const reasoning_options = existing?.reasoning_options?.length
-    ? existing.reasoning_options
-    : KiloReasoningOptions(model.opencode) ?? existing?.reasoning_options;
+  const reasoning_options = reasoning
+    ? KiloReasoningOptions(model.opencode) ?? existing?.reasoning_options
+    : undefined;
   const context = model.top_provider.context_length ?? model.context_length;
   const family = inferFamily(model, name);
   const releaseDate = dateFromTimestamp(model.created);
@@ -277,7 +270,7 @@ function KiloReasoningOptions(opencode: KiloModel["opencode"]): SyncedFullModel[
     .map(([, variant]) => variant.reasoning?.effort)
     .filter((effort): effort is string => effort !== undefined);
   const hasNone = variants.some(([, variant]) => variant.reasoning?.enabled === false);
-  const allEfforts = hasNone ? [...efforts, "none"] : [...efforts];
+  const allEfforts = [...new Set(hasNone ? [...efforts, "none"] : efforts)];
 
   if (allEfforts.length > 0) {
     const orderedEfforts = allEfforts.sort((a, b) => {
@@ -293,136 +286,3 @@ function KiloReasoningOptions(opencode: KiloModel["opencode"]): SyncedFullModel[
 
   return options.length > 0 ? options : undefined;
 }
-
-function modelMetadataExists(provider: string, modelID: string) {
-  let files = modelMetadataFilesByProvider.get(provider);
-  if (files === undefined) {
-    try {
-      files = new Set(readdirSync(path.join(MODELS_DIR, provider)));
-    } catch {
-      files = new Set();
-    }
-    modelMetadataFilesByProvider.set(provider, files);
-  }
-  return files.has(`${modelID}.toml`);
-}
-
-function baseModelOmit(
-  modelID: string,
-  limit: SyncedFullModel["limit"],
-) {
-  const metadata = modelMetadata(modelID);
-  const omit: string[] = [];
-  const baseLimit = metadata.limit;
-  if (
-    isPlainObject(baseLimit) &&
-    baseLimit.input !== undefined &&
-    limit.input === undefined &&
-    baseLimit.context !== limit.context
-  ) {
-    omit.push("limit.input");
-  }
-
-  return omit.length > 0 ? omit : undefined;
-}
-
-function baseModelOverrides(
-  modelID: string,
-  values: Partial<SyncedFullModel>,
-) {
-  const metadata = modelMetadata(modelID);
-  const result: Record<string, unknown> = {};
-
-  for (const [key, value] of Object.entries(values)) {
-    const override = inheritedOverride(value, metadata[key]);
-    if (override !== undefined) result[key] = override;
-  }
-
-  return result;
-}
-
-function inheritedOverride(value: unknown, inherited: unknown): unknown {
-  if (value === undefined) return undefined;
-  if (sameInheritedValue(value, inherited)) return undefined;
-  if (isPlainObject(value) && isPlainObject(inherited)) {
-    const overrides = Object.fromEntries(
-      Object.entries(value)
-        .map(([key, item]) => [key, inheritedOverride(item, inherited[key])])
-        .filter(([, item]) => item !== undefined),
-    );
-    return Object.keys(overrides).length > 0 ? overrides : undefined;
-  }
-  return stripUndefined(value);
-}
-
-function stripUndefined(value: unknown): unknown {
-  if (Array.isArray(value)) return value.map(stripUndefined);
-  if (isPlainObject(value)) {
-    return Object.fromEntries(
-      Object.entries(value)
-        .filter(([, item]) => item !== undefined)
-        .map(([key, item]) => [key, stripUndefined(item)]),
-    );
-  }
-  return value;
-}
-
-function sameInheritedValue(value: unknown, inherited: unknown) {
-  return stableInheritedValue(value) === stableInheritedValue(inherited);
-}
-
-function stableInheritedValue(value: unknown): string {
-  if (Array.isArray(value)) {
-    const items = value.map(stableInheritedValue);
-    const ordered = value.every((item) => item === null || typeof item !== "object")
-      ? items.sort()
-      : items;
-    return `[${ordered.join(",")}]`;
-  }
-  if (isPlainObject(value)) {
-    return `{${Object.entries(value)
-      .filter(([, item]) => item !== undefined)
-      .sort(([a], [b]) => a.localeCompare(b))
-      .map(([key, item]) => `${JSON.stringify(key)}:${stableInheritedValue(item)}`)
-      .join(",")}}`;
-  }
-  return JSON.stringify(value);
-}
-
-function isPlainObject(value: unknown): value is Record<string, unknown> {
-  return value !== null && typeof value === "object" && !Array.isArray(value);
-}
-
-function modelMetadata(modelID: string) {
-  let metadata = modelMetadataByID.get(modelID);
-  if (metadata === undefined) {
-    const filePath = path.join(MODELS_DIR, `${modelID}.toml`);
-    metadata = Bun.TOML.parse(readFileSync(filePath, "utf8")) as Record<string, unknown>;
-    modelMetadataByID.set(modelID, metadata);
-  }
-  return metadata;
-}
-
-function canonicalCandidates(provider: string, modelID: string) {
-  const candidates = [modelID];
-
-  if (provider === "anthropic") {
-    candidates.push(modelID.replace(/(claude-(?:opus|sonnet|haiku)-\d+)\.(\d+)/, "$1-$2"));
-    candidates.push(modelID.replace(/^claude-3\.5-/, "claude-3-5-"));
-  }
-
-  if (provider === "llama") {
-    candidates.push(modelID.replace(/^llama-(\d+)-(\d+)/, "llama-$1.$2"));
-    candidates.push(modelID.replace(/^llama-(4)-(maverick|scout)$/, "llama-$1-$2-17b"));
-  }
-
-  if (provider === "mistral") {
-    candidates.push(modelID.replace(/-latest$/, ""));
-  }
-
-  if (provider === "minimax") {
-    candidates.push(modelID.replace(/^minimax-m/, "MiniMax-M"));
-  }
-
-  return [...new Set(candidates)];
-}

packages/core/src/sync/providers/llmgateway.ts

@@ -2,6 +2,7 @@ import { z } from "zod";
 
 import { describeModel } from "../../describe.js";
 import { inferKimiFamily, ModelFamilyValues } from "../../family.js";
+import { ReasoningOption } from "../../schema.js";
 import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
 import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
 
@@ -11,12 +12,14 @@ const API_ENDPOINT = "https://api.llmgateway.io/v1/models";
 // canonical prefixes understood by resolveCanonicalBaseModel. Alias the few that
 // spell the lab differently. (Mirrors huggingface's CANONICAL_ORG_PREFIXES.)
 const CANONICAL_FAMILY_ALIASES: Record<string, string> = {
+  grok: "xai",
   mistral: "mistralai",
   moonshot: "moonshotai",
 };
 
 const BASE_MODEL_ALIASES: Record<string, string> = {
   "glm-5-2": "zhipuai/glm-5.2",
+  "grok-4-6": "xai/grok-4.6",
 };
 
 const Pricing = z.object({
@@ -27,6 +30,21 @@ const Pricing = z.object({
   input_cache_write: z.string().optional(),
 });
 
+const LLMGatewayProvider = z.object({
+  reasoning_efforts: z.array(z.string()).optional(),
+}).passthrough();
+
+const ReasoningEffortOrder = new Map([
+  "none",
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max",
+  "default",
+].map((effort, index) => [effort, index]));
+
 export const LLMGatewayModel = z.object({
   id: z.string(),
   name: z.string(),
@@ -37,6 +55,7 @@ export const LLMGatewayModel = z.object({
     output_modalities: z.array(z.string()),
   }),
   pricing: Pricing,
+  providers: z.array(LLMGatewayProvider),
   context_length: z.number(),
   supported_parameters: z.array(z.string()),
   structured_outputs: z.boolean().optional(),
@@ -138,12 +157,14 @@ export function buildLLMGatewayModel(
   const completion = price(model.pricing.completion);
   const reasoning = model.supported_parameters.includes("reasoning")
     || model.supported_parameters.includes("include_reasoning");
+  const reasoningOptions = llmGatewayReasoningOptions(model, existing);
   const context = model.context_length > 0
     ? model.context_length
     : existing?.limit?.context ?? model.context_length;
 
-  // The gateway is authoritative for the volatile, gateway-specific data — cost
-  // and served limits. Its supported_parameters / modalities are too noisy to
+  // The gateway is authoritative for the volatile, gateway-specific data — cost,
+  // served limits, and explicitly advertised reasoning efforts. Its
+  // supported_parameters / modalities are too noisy to
   // drive capability fields (it omits "tools" for flagship models yet lists
   // "temperature" for ones the catalog deliberately marks temperature=false),
   // so those stay curated: preserved from the existing entry (which, for a
@@ -183,6 +204,7 @@ export function buildLLMGatewayModel(
           modalities: existing.modalities,
         }),
         reasoning: existing.reasoning,
+        reasoning_options: reasoningOptions,
         temperature: existing.temperature,
         tool_call: existing.tool_call,
         structured_output: existing.structured_output,
@@ -218,6 +240,7 @@ export function buildLLMGatewayModel(
       last_updated: existing.last_updated ?? dateFromTimestamp(model.created),
       attachment: existing.attachment ?? false,
       reasoning: existing.reasoning ?? false,
+      reasoning_options: reasoningOptions,
       temperature: existing.temperature ?? false,
       tool_call: existing.tool_call ?? false,
       structured_output: existing.structured_output,
@@ -240,7 +263,11 @@ export function buildLLMGatewayModel(
   const canonical = resolveLLMGatewayBaseModel(model);
   if (canonical !== undefined) {
     const factoredLimit = { context, input: undefined, output: undefined };
-    return factorBaseModel(canonical, { limit: factoredLimit, cost }, factoredLimit);
+    return factorBaseModel(canonical, {
+      reasoning_options: reasoningOptions,
+      limit: factoredLimit,
+      cost,
+    }, factoredLimit);
   }
 
   // Brand-new model: best-effort translation from the gateway. Capability and
@@ -265,6 +292,7 @@ export function buildLLMGatewayModel(
     last_updated: dateFromTimestamp(model.created),
     attachment: input.some((value) => value !== "text"),
     reasoning,
+    reasoning_options: reasoningOptions,
     temperature: model.supported_parameters.includes("temperature"),
     tool_call: model.supported_parameters.includes("tools")
       || model.supported_parameters.includes("tool_choice"),
@@ -276,6 +304,27 @@ export function buildLLMGatewayModel(
   } satisfies SyncedFullModel;
 }
 
+function llmGatewayReasoningOptions(
+  model: LLMGatewayModel,
+  existing: ExistingModel | undefined,
+): SyncedFullModel["reasoning_options"] {
+  const advertised = new Set(model.providers.flatMap((provider) => provider.reasoning_efforts ?? []));
+  if (advertised.size === 0) return undefined;
+
+  const efforts = [...advertised].sort((a, b) => {
+    const order = (ReasoningEffortOrder.get(a) ?? Number.MAX_SAFE_INTEGER)
+      - (ReasoningEffortOrder.get(b) ?? Number.MAX_SAFE_INTEGER);
+    return order || a.localeCompare(b);
+  });
+  const preserved = existing?.reasoning_options?.filter((option) =>
+    option.type !== "effort" && !(option.type === "toggle" && advertised.has("none"))
+  ) ?? [];
+  return [
+    ...preserved,
+    ReasoningOption.parse({ type: "effort", values: efforts }),
+  ];
+}
+
 function defaultModalities(model: LLMGatewayModel) {
   return {
     input: modalities(model.architecture.input_modalities, ["text"]),

packages/core/src/sync/providers/merge-gateway.ts

@@ -13,6 +13,7 @@ const VendorReasoning = z.object({
   disable_supported: z.boolean().optional(),
   default_enabled: z.boolean().optional(),
   controls: z.array(z.string()).optional(),
+  effort_values: z.array(z.string()).optional(),
   output_style: z.string().nullable().optional(),
 }).passthrough();
 
@@ -161,6 +162,28 @@ export const mergeGateway = {
   },
 } satisfies SyncProvider<MergeGatewayModel>;
 
+export function mergeGatewayReasoningOptions(
+  reasoning: MergeGatewayVendor["capabilities"]["reasoning"],
+): NonNullable<SyncedFullModel["reasoning_options"]> | undefined {
+  if (reasoning == null) return undefined;
+  const options: NonNullable<SyncedFullModel["reasoning_options"]> = [];
+
+  if (reasoning.disable_supported === true) {
+    options.push({ type: "toggle" as const });
+  }
+
+  const controls = (reasoning.controls ?? []).map((control) => control.toLowerCase());
+  const effortValues = reasoning.effort_values ?? [];
+  if (
+    (controls.includes("reasoning.effort") || controls.includes("reasoning_effort"))
+    && effortValues.length > 0
+  ) {
+    options.push({ type: "effort" as const, values: [...effortValues] });
+  }
+
+  return options;
+}
+
 export function selectMergeGatewayVendor(model: MergeGatewayModel) {
   const canonical = model.vendors[model.provider];
   if (canonical?.availability_status === "available") {
@@ -234,8 +257,8 @@ export function buildMergeGatewayModel(
   const reasoning = routeConfirmsReasoning ? true : existing?.reasoning;
   const existingReasoningOptions = existing?.reasoning_options ?? [];
   const reasoningOptions = reasoning === true && existingReasoningOptions.length === 0
-    && selected.info.capabilities.reasoning?.disable_supported === true
-    ? [{ type: "toggle" as const }]
+    ? mergeGatewayReasoningOptions(selected.info.capabilities.reasoning)
+      ?? existingReasoningOptions
     : reasoning === true
       ? existingReasoningOptions
       : existing?.reasoning_options;

packages/core/src/sync/providers/nano-gpt.ts

@@ -51,7 +51,7 @@ export const NanoGptModel = z.object({
 }).passthrough();
 
 export const NanoGptResponse = z.object({
-  data: z.array(NanoGptModel),
+  data: z.array(NanoGptModel).min(1),
 }).passthrough();
 
 export type NanoGptModel = z.infer<typeof NanoGptModel>;
@@ -96,6 +96,7 @@ const BASE_MODEL_ALIASES: Record<string, string | undefined> = {
   "claude-opus-4": "anthropic/claude-opus-4-0",
   "claude-sonnet-4": "anthropic/claude-sonnet-4-0",
   "cohere/north-mini-code": "cohere/north-mini-code-1-0",
+  "doubao-seed-2-0-code-preview-260215": "bytedance-seed/seed-2.0-code",
 };
 
 const NANO_GPT_VARIANT_SUFFIX = /(?::(?:thinking|none|minimal|low|medium|high|xhigh|max|\d+)|-thinking)$/i;
@@ -138,8 +139,10 @@ export function buildNanoGptModel(
   const inputLimit = sourceContext ?? existing?.limit?.input;
   const outputLimit = sourceOutputLimit ?? existing?.limit?.output;
   const releaseDate = dateFromTimestamp(model.created) ?? existing?.release_date;
-  const inferredSourceReasoning = capabilities.reasoning
-    ?? (model.reasoning_efforts != null ? true : undefined);
+  const hasReasoningEfforts = model.reasoning_efforts != null && model.reasoning_efforts.length > 0;
+  const inferredSourceReasoning = hasReasoningEfforts
+    ? true
+    : capabilities.reasoning ?? (model.reasoning_efforts != null ? true : undefined);
   const reasoning = inferredSourceReasoning ?? existing?.reasoning ?? false;
   const cost = buildCost(model.pricing, existing);
   if (baseModel !== undefined) {
@@ -255,8 +258,11 @@ function reasoningOptions(
   if (reasoning === false) return undefined;
   if (reasoning === undefined) return existing;
   if (model.reasoning_efforts == null) return existing ?? [];
-  if (model.reasoning_efforts.length === 0) return [];
-  return [{ type: "effort", values: [...model.reasoning_efforts] }];
+  if (model.reasoning_efforts.length === 0) return existing ?? [];
+  const order = ReasoningEffort.options;
+  const efforts = [...new Set(model.reasoning_efforts)]
+    .sort((a, b) => order.indexOf(a) - order.indexOf(b));
+  return [{ type: "effort", values: efforts }];
 }
 
 export function resolveNanoGptBaseModel(modelID: string) {

packages/core/src/sync/providers/ofox.ts

@@ -0,0 +1,128 @@
+import { z } from "zod";
+
+import type { ExistingModel, SyncProvider, SyncedModel } from "../index.js";
+
+const API_ENDPOINT = "https://api.ofox.ai/v2/models/catalog?include=provider_price&limit=500";
+
+const Pricing = z
+  .object({
+    input: z.string().optional(),
+    output: z.string().optional(),
+    input_cache_read: z.string().optional(),
+    input_cache_write: z.string().optional(),
+    input_cache_write_5m: z.string().optional(),
+    input_cache_write_1h: z.string().optional(),
+  })
+  .passthrough();
+
+const ProviderPrice = z
+  .object({
+    pricing: Pricing.optional(),
+    is_override: z.boolean().optional(),
+  })
+  .passthrough();
+
+export const OfoxModel = z
+  .object({
+    id: z.string().min(1),
+    display_name: z.string().optional(),
+    mode: z.string(),
+    context_window: z.number().optional(),
+    max_output_tokens: z.number().optional(),
+    pricing: Pricing.optional(),
+    provider_price: ProviderPrice.nullable().optional(),
+    is_deprecated: z.boolean().optional(),
+  })
+  .passthrough();
+
+export const OfoxResponse = z
+  .object({
+    data: z.array(OfoxModel),
+  })
+  .passthrough();
+
+export type OfoxModel = z.infer<typeof OfoxModel>;
+
+/**
+ * Ofox lists a curated subset of its catalog here, so this sync only updates
+ * existing TOMLs (`skipCreates`) and treats the Ofox catalog API as
+ * authoritative for pricing and deprecation status only. Everything else in
+ * the authored TOMLs — `base_model` inheritance, `reasoning_options`, and the
+ * per-model `[provider]` protocol overrides — is preserved as hand-authored.
+ */
+export const ofox = {
+  id: "ofox",
+  name: "Ofox",
+  modelsDir: "providers/ofox/models",
+  skipCreates: true,
+  trackMissingModels: true,
+  deleteMissing: false,
+  sourceID(model) {
+    return model.mode === "chat" ? model.id : undefined;
+  },
+  missingNotice(paths) {
+    return paths.map(
+      (file) => `Ofox catalog no longer lists ${file}; review for manual deprecation or removal.`,
+    );
+  },
+  async fetchModels() {
+    const response = await fetch(API_ENDPOINT);
+    if (!response.ok) {
+      throw new Error(`Ofox request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    return OfoxResponse.parse(raw).data;
+  },
+  translateModel(model, context) {
+    if (model.mode !== "chat") return undefined;
+    const authored = context.authored(model.id);
+    if (authored === undefined) return undefined;
+    return {
+      id: model.id,
+      model: buildOfoxModel(model, authored),
+    };
+  },
+} satisfies SyncProvider<OfoxModel>;
+
+/** Ofox prices are $/token decimal strings; catalog omits zero-value fields. */
+function price(value: string | undefined) {
+  if (value === undefined) return undefined;
+  const number = Number(value);
+  if (!Number.isFinite(number) || number <= 0) return undefined;
+  return Math.round(number * 1_000_000_000_000) / 1_000_000;
+}
+
+export function buildOfoxModel(model: OfoxModel, authored: ExistingModel): SyncedModel {
+  const { id: _authoredID, ...preserved } = authored;
+  // Effective customer price: when ops set a provider_price override (price
+  // cuts / promos), that is what users are billed; the base `pricing` keeps
+  // the pre-discount list price.
+  const pricing =
+    model.provider_price?.is_override === true && model.provider_price.pricing !== undefined
+      ? model.provider_price.pricing
+      : model.pricing;
+  const input = price(pricing?.input);
+  const output = price(pricing?.output);
+  const cacheRead = price(pricing?.input_cache_read);
+  const cacheWrite = price(pricing?.input_cache_write) ?? price(pricing?.input_cache_write_5m);
+  const cost =
+    input !== undefined || output !== undefined
+      ? {
+          ...authored.cost,
+          input: input ?? 0,
+          output: output ?? 0,
+          cache_read: cacheRead,
+          cache_write: cacheWrite,
+        }
+      : authored.cost;
+
+  const status = model.is_deprecated === true ? ("deprecated" as const) : authored.status;
+
+  return {
+    ...preserved,
+    cost,
+    status,
+  } as SyncedModel;
+}

packages/core/src/sync/providers/openrouter.ts

@@ -13,6 +13,7 @@ const modelMetadataFilesByProvider = new Map<string, Set<string>>();
 let allModelMetadataIDs: string[] | undefined;
 
 const CANONICAL_BASE_MODEL_OVERRIDES = {
+  "bytedance/dola-seed-2.0-code": "bytedance-seed/seed-2.0-code",
   "openai/gpt-5.6-luna-pro": "openai/gpt-5.6-luna",
   "openai/gpt-5.6-sol-pro": "openai/gpt-5.6-sol",
   "openai/gpt-5.6-terra-pro": "openai/gpt-5.6-terra",
@@ -23,6 +24,7 @@ const CANONICAL_BASE_MODEL_OVERRIDES = {
 const CANONICAL_PROVIDER_PREFIXES = {
   alibaba: { provider: "alibaba", metadata: "alibaba" },
   anthropic: { provider: "anthropic", metadata: "anthropic" },
+  "bytedance-seed": { provider: "bytedance-seed", metadata: "bytedance-seed" },
   cohere: { provider: "cohere", metadata: "cohere" },
   deepseek: { provider: "deepseek", metadata: "deepseek" },
   google: { provider: "google", metadata: "google" },
@@ -319,6 +321,9 @@ function openRouterReasoningOptions(reasoning: OpenRouterModel["reasoning"]): Sy
     : reasoning.supported_efforts;
 
   if (efforts !== undefined) {
+    if (!reasoning.mandatory && !efforts.includes("none")) {
+      options.push({ type: "toggle" });
+    }
     options.push({
       type: "effort",
       values: reasoning.mandatory ? efforts.filter((value) => value !== "none") : [...efforts],
@@ -541,7 +546,7 @@ function isPlainObject(value: unknown): value is Record<string, unknown> {
   return value !== null && typeof value === "object" && !Array.isArray(value);
 }
 
-function modelMetadata(modelID: string) {
+export function modelMetadata(modelID: string) {
   let metadata = modelMetadataByID.get(modelID);
   if (metadata === undefined) {
     const filePath = path.join(MODELS_DIR, `${modelID}.toml`);