chore(sync): update Vercel AI Gateway model catalog

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

.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-vl-235b-a22b-instruct.toml

@@ -0,0 +1,24 @@
+name = "Qwen3 VL 235B A22B Instruct"
+description = "Qwen vision-language instruct model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2025-09-23"
+last_updated = "2025-09-23"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-03-31"
+open_weights = true
+
+[limit]
+context = 131_072
+output = 32_768
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct"

models/alibaba/qwen3-vl-235b-a22b-thinking.toml

@@ -0,0 +1,24 @@
+name = "Qwen3 VL 235B A22B Thinking"
+description = "Qwen vision-language thinking model for visual reasoning, documents, and agent tasks"
+family = "qwen"
+release_date = "2025-09-23"
+last_updated = "2025-09-23"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-03-31"
+open_weights = true
+
+[limit]
+context = 131_072
+output = 32_768
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Thinking"

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/qwen3.8-2.4t-a95b.toml

@@ -0,0 +1,33 @@
+# Sources (accessed 2026-08-16):
+# https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
+# https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/README.md
+# https://qwen.ai/blog?id=qwen3.8
+# https://openrouter.ai/qwen/qwen3.8-2.4t-a95b
+# Open-weight twin of Qwen3.8 Max: text-only, thinking always on,
+# reasoning_effort low|medium|xhigh (default xhigh). Native context 262K,
+# extensible to ~1.01M. Distinct from closed multimodal qwen3.8-max.
+
+name = "Qwen3.8 2.4T A95B"
+description = "Open-weight sparse MoE (2.4T total, 95B active), the open-weight twin of Qwen3.8 Max for coding, research, complex reasoning, and agentic workflows"
+family = "qwen"
+release_date = "2026-08-12"
+last_updated = "2026-08-12"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = true
+license = "qwen3.8-max"
+
+[limit]
+context = 262_144
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B"

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.toml

@@ -1,6 +1,10 @@
-# Sources (accessed 2026-08-03):
+# 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
@@ -9,6 +13,10 @@
 # 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"
@@ -26,5 +34,5 @@ context = 1_000_000
 output = 131_072
 
 [modalities]
-input = ["text", "image", "video"]
+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/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-vision-exp.toml

@@ -0,0 +1,19 @@
+name = "DeepSeek V4 Flash Vision Exp"
+description = "Experimental multimodal DeepSeek V4 Flash model for image understanding, coding, and agentic work"
+family = "deepseek-flash"
+release_date = "2026-08-21"
+last_updated = "2026-08-21"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/deepseek/deepseek-v4-pro-0423.toml

@@ -0,0 +1,22 @@
+# https://ofox.ai/models/deepseek/deepseek-v4-pro-0423
+# https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
+name = "DeepSeek V4 Pro 0423"
+description = "DeepSeek V4 Pro initial snapshot with million-token context and support for thinking and non-thinking modes"
+family = "deepseek-thinking"
+release_date = "2026-04-23"
+last_updated = "2026-04-23"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-05"
+open_weights = true
+
+[limit]
+context = 1_000_000
+output = 384_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

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-3.7-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/google/gemini-flash-latest.toml

@@ -1,15 +1,15 @@
-# Tracks the current Gemini Flash release (gemini-3.5-flash).
+# Tracks the current Gemini Flash release (gemini-3.7-flash).
 name = "Gemini Flash Latest"
-description = "Fast Gemini model balancing multimodal reasoning, tool use, and cost"
+description = "High-efficiency Gemini model for agentic workflows, coding, and multimodal reasoning"
 family = "gemini-flash"
-release_date = "2026-05-19"
-last_updated = "2026-05-19"
+release_date = "2026-08-13"
+last_updated = "2026-08-13"
 attachment = true
 reasoning = true
 temperature = true
 tool_call = true
 structured_output = true
-knowledge = "2025-01"
+knowledge = "2026-03"
 open_weights = false
 
 [limit]

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

@@ -1,15 +1,15 @@
-# Tracks the current Gemini Flash-Lite release (gemini-3.1-flash-lite).
+# Tracks the current Gemini Flash-Lite release (gemini-3.5-flash-lite).
 name = "Gemini Flash-Lite Latest"
-description = "Low-latency Gemini model for high-volume multimodal and agent workloads"
+description = "Fast Gemini model balancing multimodal reasoning, tool use, and cost"
 family = "gemini-flash-lite"
-release_date = "2026-05-07"
-last_updated = "2026-05-07"
+release_date = "2026-07-21"
+last_updated = "2026-07-21"
 attachment = true
 reasoning = true
 temperature = true
 tool_call = true
 structured_output = true
-knowledge = "2025-01"
+knowledge = "2026-03"
 open_weights = false
 
 [limit]

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-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/mistral/voxtral-small-latest.toml

@@ -0,0 +1,24 @@
+# Sources (accessed 2026-08-16):
+#   https://docs.mistral.ai/models/model-cards/voxtral-small-25-07
+#   https://mistral.ai/news/voxtral/
+# Field values mirror Mistral's own first-party host entry in this repo
+# (providers/mistral/models/voxtral-small-latest.toml); host-scoped keys
+# (cost, status) are intentionally left to the provider files.
+name = "Voxtral Small (latest)"
+description = "Instruct model with native audio input for speech understanding and tool use"
+family = "voxtral"
+release_date = "2025-07-15"
+last_updated = "2025-07-15"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 32_000
+output = 32_000
+
+[modalities]
+input = ["text", "audio"]
+output = ["text"]

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/perplexity/sonar-deep-research.toml

@@ -0,0 +1,25 @@
+# Sources (accessed 2026-08-16):
+#   https://docs.perplexity.ai/docs/sonar/models/sonar-deep-research
+#   https://docs.perplexity.ai/api-reference/sonar-post
+# Field values mirror Perplexity's own first-party host entry in this repo
+# (providers/perplexity/models/sonar-deep-research.toml); host-scoped keys
+# (cost, reasoning_options) are intentionally left to the provider files.
+name = "Sonar Deep Research"
+description = "Sonar search model for autonomous research and citation-backed long-form reports"
+family = "sonar"
+release_date = "2025-02-01"
+last_updated = "2025-09-01"
+attachment = false
+reasoning = true
+temperature = false
+tool_call = false
+knowledge = "2025-01"
+open_weights = false
+
+[limit]
+context = 128_000
+output = 32_768
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/sakana/sakana-namazu.toml

@@ -0,0 +1,59 @@
+name = "Sakana Namazu"
+description = "Japanese-specialized reasoning model based on Kimi K2.6 and tuned for Japanese language, culture, and business workflows"
+family = "sakana-namazu"
+release_date = "2026-08-03"
+last_updated = "2026-08-03"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 262_144
+output = 65_536
+
+[modalities]
+input = ["text", "image", "pdf"]
+output = ["text"]
+
+[[links]]
+label = "Official product page"
+url = "https://sakana.ai/namazu/"
+type = "announcement"
+
+[[links]]
+label = "Official model documentation"
+url = "https://console.sakana.ai/models?model=sakana-namazu"
+type = "docs"
+
+[[benchmarks]]
+name = "AIME26"
+score = 96.67
+source = "https://console.sakana.ai/models?model=sakana-namazu"
+
+[[benchmarks]]
+name = "MMLU-Pro"
+score = 90.33
+source = "https://console.sakana.ai/models?model=sakana-namazu"
+
+[[benchmarks]]
+name = "LiveCodeBench v6"
+score = 90.33
+source = "https://console.sakana.ai/models?model=sakana-namazu"
+
+[[benchmarks]]
+name = "JFBench"
+score = 37.40
+source = "https://console.sakana.ai/models?model=sakana-namazu"
+
+[[benchmarks]]
+name = "Translation"
+score = 52.20
+source = "https://console.sakana.ai/models?model=sakana-namazu"
+
+[[benchmarks]]
+name = "FairPoliticsQA"
+score = 56.30
+source = "https://console.sakana.ai/models?model=sakana-namazu"

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/swiss-ai/apertus-8b.toml

@@ -0,0 +1,33 @@
+# Sources (accessed 2026-08-16):
+#   https://huggingface.co/swiss-ai/Apertus-8B-Instruct-2509
+#   https://arxiv.org/abs/2509.14233
+# The model card states "Apertus by default supports a context length up to 65,536 tokens",
+# Apache-2.0 licensing, and tool use support. Sibling entry: models/swiss-ai/apertus-70b.toml.
+name = "Apertus 8B"
+description = "Fully open 8B multilingual LLM supporting 1800+ languages with 65K context. Trained on 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-8B-Instruct-2509"
+
+[[links]]
+label = "Paper"
+url = "https://arxiv.org/abs/2509.14233"
+type = "paper"

models/trendyol/asure-12b.toml

@@ -0,0 +1,30 @@
+# Sources:
+# https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B
+# https://huggingface.co/api/models/Trendyol/Trendyol-LLM-Asure-12B (createdAt, license, base_model)
+# https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B/raw/main/config.json (max_position_embeddings)
+# `reasoning` and `tool_call` are not stated on the model card; both were
+# measured against a host serving these weights (llmtr.com, 2026-08-16):
+# a request carrying `tools` returns no tool_calls, and no reasoning output
+# is produced.
+name = "Trendyol Asure 12B"
+description = "Turkish-language multimodal instruct model built on Gemma 3 12B for e-commerce text, chat, and image-text tasks"
+family = "gemma"
+release_date = "2026-02-19"
+last_updated = "2026-02-20"
+attachment = true
+reasoning = false
+temperature = true
+tool_call = false
+open_weights = true
+license = "Gemma"
+
+[limit]
+context = 131_072
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/Trendyol/Trendyol-LLM-Asure-12B"

models/upstage/solar-pro2.toml

@@ -0,0 +1,25 @@
+# Sources (accessed 2026-08-16):
+#   https://developers.upstage.ai/docs/apis/chat
+#   https://developers.upstage.ai/docs/capabilities/generate/reasoning
+# Field values mirror Upstage's own first-party host entry in this repo
+# (providers/upstage/models/solar-pro2.toml); host-scoped keys (cost,
+# reasoning_options) are intentionally left to the provider files.
+name = "Solar Pro 2"
+description = "Flagship model for demanding analysis, coding, and production agent workflows"
+family = "solar-pro"
+release_date = "2025-05-20"
+last_updated = "2025-05-20"
+attachment = false
+reasoning = true
+temperature = true
+knowledge = "2025-03"
+tool_call = true
+open_weights = false
+
+[limit]
+context = 65_536
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/upstage/solar-pro3.toml

@@ -0,0 +1,25 @@
+# Sources (accessed 2026-08-16):
+#   https://developers.upstage.ai/docs/apis/chat
+#   https://developers.upstage.ai/docs/capabilities/generate/reasoning
+# Field values mirror Upstage's own first-party host entry in this repo
+# (providers/upstage/models/solar-pro3.toml); host-scoped keys (cost,
+# reasoning_options) are intentionally left to the provider files.
+name = "Solar Pro 3"
+description = "Flagship model for demanding analysis, coding, and production agent workflows"
+family = "solar-pro"
+release_date = "2026-01"
+last_updated = "2026-01"
+attachment = false
+reasoning = true
+temperature = true
+knowledge = "2025-03"
+tool_call = true
+open_weights = false
+
+[limit]
+context = 131_072
+output = 8_192
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/upstage/solar-pro4.toml

@@ -0,0 +1,26 @@
+# Sources (accessed 2026-08-16):
+#   https://developers.upstage.ai/docs/apis/chat
+#   https://developers.upstage.ai/docs/capabilities/generate/reasoning
+# Field values mirror Upstage's own first-party host entry in this repo
+# (providers/upstage/models/solar-pro4.toml); host-scoped keys (cost,
+# reasoning_options) are intentionally left to the provider files.
+name = "Solar Pro 4"
+description = "Upstage's flagship model, specialized for agentic use"
+family = "solar-pro"
+release_date = "2026-08-06"
+last_updated = "2026-08-06"
+attachment = false
+reasoning = true
+temperature = true
+knowledge = "2026-02"
+tool_call = true
+structured_output = true
+open_weights = false
+
+[limit]
+context = 524_288
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]

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"

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/xai/grok-imagine-image-2.0.toml

@@ -0,0 +1,26 @@
+# Sources:
+# - https://docs.x.ai/docs/models
+# - https://docs.x.ai/developers/models/grok-imagine-image-2.0
+# - https://docs.x.ai/docs/guides/image-generation
+# - https://x.ai/news/grok-imagine-image-2
+# Pricing: $0.04 per image (not token-based; no [cost] authored)
+# Release: 2026-08-07 (GA as Quality Mode; API model id grok-imagine-image-2.0)
+
+name = "Grok Imagine Image 2.0"
+description = "Image model for prompt-driven generation, editing, and visual design workflows"
+family = "grok"
+release_date = "2026-08-07"
+last_updated = "2026-08-07"
+attachment = true
+reasoning = false
+temperature = false
+tool_call = false
+open_weights = false
+
+[limit]
+context = 8_000
+output = 0
+
+[modalities]
+input = ["text", "image"]
+output = ["image"]

models/zhipuai/glm-4.6v-flash.toml

@@ -0,0 +1,25 @@
+# Sources (accessed 2026-08-19):
+# - https://z.ai/blog/glm-4.6v
+# - https://huggingface.co/zai-org/GLM-4.6V-Flash
+name = "GLM-4.6V-Flash"
+description = "Lightweight GLM vision model for visual reasoning, documents, and multimodal agents"
+family = "glm"
+release_date = "2025-12-08"
+last_updated = "2025-12-08"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 128_000
+output = 32_768
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/zai-org/GLM-4.6V-Flash"

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 +28,8 @@
     "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",
+    "llmgateway-providers:sync": "bun ./packages/core/script/sync-models.ts llmgateway-providers",
+    "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",

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

@@ -11,7 +11,16 @@ const diff = Bun.spawnSync(["git", "diff", "--name-status", "--no-renames", base
 
 if (diff.exitCode !== 0) process.exit(diff.exitCode ?? 1);
 
-const decision = await classifyAutoMerge(parseNameStatus(diff.stdout.toString()));
+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("; ")}.`;

packages/core/src/family.ts

@@ -30,6 +30,7 @@ export const ModelFamilyValues = [
   "claude-sonnet",
   "claude-opus",
   "claude-fable",
+  "claude-mythos",
 
   // Gemini style
   "gemini",
@@ -51,6 +52,7 @@ export const ModelFamilyValues = [
 
   // Meta Muse
   "muse",
+  "muse-free",
 
   // Alibaba Qwen
   "qwen",
@@ -365,6 +367,9 @@ export const ModelFamilyValues = [
   // Conductor
   "fugu",
 
+  // Sakana Namazu
+  "sakana-namazu",
+
   // V0
   "v0",
 

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

@@ -1,9 +1,22 @@
 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(["openrouter"]);
+const REVIEWED_REASONING_PROVIDERS = new Set([
+  "crossmodel",
+  "edenai",
+  "empiriolabs",
+  "hyper",
+  "kilo",
+  "llmgateway",
+  "llmgateway-providers",
+  "merge-gateway",
+  "nano-gpt",
+  "openrouter",
+  "venice",
+]);
 
 export interface CatalogChange {
   status: "created" | "updated" | "deleted";
@@ -29,6 +42,7 @@ function isProviderModel(path: string) {
 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;
@@ -42,18 +56,34 @@ export async function classifyAutoMerge(
     reasons.push(`${created + deleted} models created or deleted (limit ${MAX_MODEL_CHURN})`);
   }
 
-  for (const change of models) {
-    if (change.status === "deleted" || !isProviderModel(change.path)) continue;
-
-    const model = Bun.TOML.parse(await load(change.path)) as Record<string, unknown>;
+  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 load(`models/${model.base_model}.toml`)) as Record<string, unknown>;
+      const base = Bun.TOML.parse(await loader(`models/${model.base_model}.toml`)) as Record<string, unknown>;
       reasoning = base.reasoning;
     }
 
-    if (reasoning === true) {
-      if (!Object.hasOwn(model, "reasoning_options")) {
+    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`);

packages/core/src/sync/index.ts

@@ -10,15 +10,18 @@ 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 { llmgateway, llmgatewayProviders } from "./providers/llmgateway.js";
 import { mergeGateway } from "./providers/merge-gateway.js";
 import { nanoGpt } from "./providers/nano-gpt.js";
 import { openai } from "./providers/openai.js";
@@ -26,6 +29,7 @@ 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";
@@ -94,7 +98,17 @@ export interface SyncProvider<SourceModel> {
       existing(id: string): ExistingModel | undefined;
       authored(id: string): ExistingModel | undefined;
     },
-  ): { id: string; model: SyncedModel; metadata?: { id: string; model: SyncedMetadata } } | undefined;
+  ): {
+    id: string;
+    model: SyncedModel;
+    metadata?: { id: string; model: SyncedMetadata };
+    /**
+     * Leading comment block for the written file when it has none of its own
+     * (e.g. the wire-path header every toggle reasoning control requires). A
+     * header already present on the existing file always wins.
+     */
+    header?: string;
+  } | undefined;
 }
 
 export interface SyncResult {
@@ -115,15 +129,19 @@ 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>;
+  "llmgateway-providers": SyncProvider<any>;
   "merge-gateway": SyncProvider<any>;
   "nano-gpt": SyncProvider<any>;
   ofox: SyncProvider<any>;
@@ -131,6 +149,7 @@ export const providers: {
   openrouter: SyncProvider<any>;
   ovhcloud: SyncProvider<any>;
   pioneer: SyncProvider<any>;
+  requesty: SyncProvider<any>;
   tinfoil: SyncProvider<any>;
   vercel: SyncProvider<any>;
   venice: SyncProvider<any>;
@@ -142,15 +161,19 @@ export const providers: {
   baseten,
   chutes,
   "cloudflare-workers-ai": cloudflareWorkersAi,
+  cortecs,
   crossmodel,
   deepinfra,
   digitalocean,
+  edenai,
   empiriolabs,
   google,
   hyper,
   huggingface,
+  inceptron,
   kilo,
   llmgateway,
+  "llmgateway-providers": llmgatewayProviders,
   "merge-gateway": mergeGateway,
   "nano-gpt": nanoGpt,
   ofox,
@@ -158,6 +181,7 @@ export const providers: {
   openrouter,
   ovhcloud,
   pioneer,
+  requesty,
   tinfoil,
   vercel,
   venice,
@@ -168,18 +192,22 @@ export const providers: {
 export const groups = {
   aggregators: [
     "crossmodel",
+    "edenai",
     "empiriolabs",
     "huggingface",
+    "inceptron",
     "kilo",
     "llmgateway",
+    "llmgateway-providers",
     "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;
@@ -255,13 +283,16 @@ export async function syncProvider<SourceModel>(
       : preserveBaseModel(translated.model, existing.get(relativePath)?.authored);
     const translatedBase = "base_model" in translatedModel ? translatedModel.base_model : undefined;
     let resolvedReasoning: boolean | undefined;
+    let baseReasoningOptions: unknown;
     if (translatedBase !== undefined) {
       if (translated.metadata?.id === translatedBase) {
         resolvedReasoning = translated.metadata.model.reasoning;
+        baseReasoningOptions = translated.metadata.model.reasoning_options;
       } else {
         modelMetadata ??= await readModelMetadata(provider.modelsDir);
         const canonicalReasoning = modelMetadata[translatedBase]?.reasoning;
         resolvedReasoning = typeof canonicalReasoning === "boolean" ? canonicalReasoning : undefined;
+        baseReasoningOptions = modelMetadata[translatedBase]?.reasoning_options;
       }
     } else {
       resolvedReasoning = existing.get(relativePath)?.toml.reasoning;
@@ -270,6 +301,7 @@ export async function syncProvider<SourceModel>(
       translatedModel,
       existing.get(relativePath)?.authored,
       resolvedReasoning,
+      baseReasoningOptions,
     );
     const withDescription = provider.preserveDescriptions === false
       ? withReasoningOptions
@@ -285,7 +317,7 @@ export async function syncProvider<SourceModel>(
 
     desired.set(relativePath, {
       model: parsed.data,
-      content: (existing.get(relativePath)?.header ?? "") + formatToml(parsed.data),
+      content: ((existing.get(relativePath)?.header || translated.header) ?? "") + formatToml(parsed.data),
     });
   }
 
@@ -461,6 +493,7 @@ export function preserveReasoningOptions(
   model: SyncedModel,
   existing: ExistingModel | undefined,
   resolvedReasoning: boolean | undefined = existing?.reasoning,
+  baseReasoningOptions: unknown = undefined,
 ): SyncedModel {
   if ((model.reasoning ?? resolvedReasoning) === false) {
     const { reasoning_options: _reasoningOptions, ...withoutReasoningOptions } = model;
@@ -468,7 +501,10 @@ export function preserveReasoningOptions(
   }
   if (model.reasoning_options !== undefined) return model;
   if (existing?.reasoning_options === undefined) {
-    return (model.reasoning ?? resolvedReasoning) === true
+    // When the base model already declares reasoning_options, leave the field
+    // unset so the factored file inherits them — stamping [] here would
+    // shadow the base's real controls with "no controls".
+    return (model.reasoning ?? resolvedReasoning) === true && baseReasoningOptions === undefined
       ? { ...model, reasoning_options: [] }
       : model;
   }

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

@@ -1,7 +1,13 @@
 import { z } from "zod";
 
 import { describeModel } from "../../describe.js";
-import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import type {
+  ExistingModel,
+  SyncProvider,
+  SyncedBaseModel,
+  SyncedFullModel,
+  SyncedModel,
+} from "../index.js";
 import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
 
 const API_ENDPOINT = "https://inference.baseten.co/v1/models";
@@ -61,6 +67,7 @@ export const baseten = {
   },
   translateModel(model, context) {
     const existing = context.existing(model.id);
+    const authored = context.authored(model.id);
     const baseModel = existing === undefined
       ? resolveBasetenBaseModel(model.id)
       : existing.base_model;
@@ -72,7 +79,7 @@ export const baseten = {
 
     return {
       id: model.id,
-      model: buildBasetenModel(model, existing, baseModel),
+      model: buildBasetenModel(model, existing, baseModel, authored),
     };
   },
 } satisfies SyncProvider<BasetenModel>;
@@ -102,6 +109,7 @@ export function buildBasetenModel(
   model: BasetenModel,
   existing: ExistingModel | undefined,
   baseModel = existing === undefined ? resolveBasetenBaseModel(model.id) : existing.base_model,
+  authored?: ExistingModel,
 ): SyncedModel {
   const features = new Set(model.supported_features);
   const samplingParameters = new Set(model.supported_sampling_parameters);
@@ -122,7 +130,9 @@ export function buildBasetenModel(
   const limit = {
     context: model.context_length,
     input: existing?.limit?.input,
-    output: model.max_completion_tokens,
+    // Explicit provider limits are serving overrides and sync pins. Baseten's
+    // catalog has returned a model's context window as its completion limit.
+    output: authored?.limit?.output ?? model.max_completion_tokens,
   };
   const values: Partial<SyncedFullModel> = {
     name: model.name ?? existing?.name,
@@ -159,7 +169,21 @@ export function buildBasetenModel(
     if (limit.context === undefined || limit.output === undefined) {
       throw new Error(`Baseten model ${model.id} has incomplete token limits required for sync`);
     }
-    return factorBaseModel(baseModel, values, limit, existing?.base_model_omit);
+    const factored = factorBaseModel(
+      baseModel,
+      values,
+      limit,
+      existing?.base_model_omit,
+    ) as SyncedBaseModel;
+    if (authored?.limit?.output === undefined) return factored;
+
+    return {
+      ...factored,
+      limit: {
+        ...factored.limit,
+        output: authored.limit.output,
+      },
+    };
   }
 
   const required = z.object({

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

@@ -0,0 +1,179 @@
+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"]);
+type CortecsModality = z.infer<typeof CortecsModality>;
+
+function modalities(values: string[]): CortecsModality[] {
+  const allowed = new Set<CortecsModality>(CortecsModality.options);
+  const result = values
+    .map((value) => value.toLowerCase())
+    .map((value) => value === "file" ? "pdf" : value)
+    .filter((value): value is CortecsModality => allowed.has(value as CortecsModality));
+  return [...new Set<CortecsModality>(result.length > 0 ? result : ["text"])];
+}
+
+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(z.string()).transform(modalities).default(["text"]),
+  output_modalities: z.array(z.string()).transform(modalities).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

@@ -22,13 +22,17 @@ function baseModelExists(modelID: string): boolean {
 // CROSSMODEL_MODELS_URL overrides the endpoint (e.g. a local backend) for testing.
 const API_ENDPOINT = process.env.CROSSMODEL_MODELS_URL ?? "https://www.crossmodel.ai/api/models";
 
+const REASONING_EFFORTS = ["none", "minimal", "low", "medium", "high", "xhigh", "max", "default"] as const;
+
 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.enum(REASONING_EFFORTS)).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 +59,10 @@ 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(),
@@ -155,27 +162,17 @@ function modalities(values: string[] | undefined, fallback: Modality[]): Modalit
   return [...new Set(result.length > 0 ? result : fallback)];
 }
 
-// models.dev's reasoning_options effort enum (schema.ts ReasoningEffortValue).
-// Guarding against it means an unexpected upstream value is dropped instead of
-// silently producing a TOML that fails `validate`.
-const REASONING_EFFORTS = ["none", "minimal", "low", "medium", "high", "xhigh", "max", "default"] as const;
-type ReasoningEffort = (typeof REASONING_EFFORTS)[number];
-function isReasoningEffort(value: string): value is ReasoningEffort {
-  return (REASONING_EFFORTS as readonly string[]).includes(value);
-}
-
 // Project CrossModel's capabilities.reasoning onto models.dev reasoning_options.
 //   reasoning absent  -> undefined (non-reasoning model; option omitted)
 //   reasoning === {}  -> [] (model reasons, no verified user-selectable control)
 //   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) {
-    const values = reasoning.effort.filter(isReasoningEffort);
-    if (values.length > 0) options.push({ type: "effort", values });
+    if (reasoning.effort.length > 0) options.push({ type: "effort", values: reasoning.effort });
   }
   if (reasoning.budget_tokens !== undefined) {
     const budget: { type: "budget_tokens"; min?: number; max?: number } = { type: "budget_tokens" };
@@ -186,7 +183,7 @@ function reasoningOptions(model: CrossModelModel): SyncedModel["reasoning_option
   return options;
 }
 
-function buildCrossModel(
+export function buildCrossModel(
   model: CrossModelModel,
   existing: ExistingModel | undefined,
 ): SyncedModel | undefined {
@@ -247,7 +244,7 @@ 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/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-0731",
-  "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/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/llmgateway.ts

@@ -1,22 +1,30 @@
 import { z } from "zod";
+import { existsSync, readFileSync } from "node:fs";
+import path from "node:path";
 
 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";
+import { factorBaseModel, resolveModelMetadataBaseModel } from "./openrouter.js";
 
 const API_ENDPOINT = "https://api.llmgateway.io/v1/models";
 
 // LLM Gateway names the originating lab in `family`; most already match the
-// canonical prefixes understood by resolveCanonicalBaseModel. Alias the few that
-// spell the lab differently. (Mirrors huggingface's CANONICAL_ORG_PREFIXES.)
+// canonical prefixes understood by resolveModelMetadataBaseModel, and labs
+// outside that shared table (e.g. perplexity) resolve through its exact
+// `models/` path match without widening the OpenRouter prefix map for every
+// other provider. 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 +35,17 @@ const Pricing = z.object({
   input_cache_write: z.string().optional(),
 });
 
+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(),
@@ -36,8 +55,22 @@ export const LLMGatewayModel = z.object({
     input_modalities: z.array(z.string()),
     output_modalities: z.array(z.string()),
   }),
+  providers: z.array(
+    z.object({
+      providerId: z.string().optional(),
+      vision: z.boolean().optional(),
+      tools: z.boolean().optional(),
+      reasoning: z.boolean().optional(),
+      reasoning_efforts: z.array(
+        z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max", "default"]),
+      ).optional(),
+    }).passthrough(),
+  ).optional(),
   pricing: Pricing,
-  context_length: z.number(),
+  // Absent for pseudo-models (custom/auto) and some non-text mappings; text
+  // models always report it.
+  context_length: z.number().optional(),
+  max_output: z.number().optional(),
   supported_parameters: z.array(z.string()),
   structured_outputs: z.boolean().optional(),
 }).passthrough();
@@ -48,31 +81,117 @@ export const LLMGatewayResponse = z.object({
 
 export type LLMGatewayModel = z.infer<typeof LLMGatewayModel>;
 
+async function fetchLLMGatewayModels(url: string) {
+  const headers = process.env.LLMGATEWAY_API_KEY
+    ? { Authorization: `Bearer ${process.env.LLMGATEWAY_API_KEY}` }
+    : undefined;
+  const response = await fetch(url, { headers });
+  if (!response.ok) {
+    throw new Error(`LLM Gateway request failed: ${response.status} ${response.statusText}`);
+  }
+  return response.json();
+}
+
+function textOnly(model: LLMGatewayModel) {
+  const output = model.architecture.output_modalities;
+  return output.length === 1 && output[0] === "text";
+}
+
+// The DevPass (LLM Gateway) provider: the gateway's aggregated catalog of root
+// model IDs, auto-routed across upstream providers.
 export const llmgateway = {
   id: "llmgateway",
-  name: "LLM Gateway",
+  name: "DevPass (LLM Gateway)",
   modelsDir: "providers/llmgateway/models",
   async fetchModels() {
-    const headers = process.env.LLMGATEWAY_API_KEY
-      ? { Authorization: `Bearer ${process.env.LLMGATEWAY_API_KEY}` }
-      : undefined;
-    const response = await fetch(API_ENDPOINT, { headers });
-    if (!response.ok) {
-      throw new Error(`LLM Gateway request failed: ${response.status} ${response.statusText}`);
-    }
-    return response.json();
+    return fetchLLMGatewayModels(API_ENDPOINT);
   },
   parseModels(raw) {
-    return LLMGatewayResponse.parse(raw).data.filter((model) => {
-      const output = model.architecture.output_modalities;
-      return output.length === 1 && output[0] === "text";
-    });
+    const data = LLMGatewayResponse.parse(raw).data.filter(textOnly);
+    // An empty catalog is an upstream fault; syncing it would delete every
+    // model file, so fail loudly instead.
+    if (data.length === 0) {
+      throw new Error("LLM Gateway returned no text models");
+    }
+    return data;
   },
   translateModel(model, context) {
-    return {
-      id: model.id,
-      model: buildLLMGatewayModel(model, context.existing(model.id)),
-    };
+    const translated = buildLLMGatewayModel(model, context.existing(model.id));
+    if (translated === undefined) {
+      return undefined;
+    }
+    return { id: model.id, model: translated };
+  },
+  sourceID(model) {
+    return model.id;
+  },
+} satisfies SyncProvider<LLMGatewayModel>;
+
+// Every toggle reasoning control requires a leading wire-path comment, and the
+// sync runner only carries over headers that already exist on disk. Files this
+// sync writes with a toggle get the gateway-wide default; a hand-written
+// header on the existing file always wins.
+const TOGGLE_HEADER = `# Toggle: $.reasoning_effort = "none" disables thinking; any other accepted
+# value (or omitting the field) leaves it on. The gateway maps it to the
+# deployment's thinking switch.
+# https://docs.llmgateway.io/features/reasoning
+`;
+
+function toggleHeader(model: SyncedModel) {
+  return model.reasoning_options?.some((option) => option.type === "toggle")
+    ? TOGGLE_HEADER
+    : undefined;
+}
+
+// The LLM Gateway provider: one entry per upstream provider mapping, addressed
+// the way the gateway accepts provider-pinned requests (`provider/model-id`).
+export const llmgatewayProviders = {
+  id: "llmgateway-providers",
+  name: "LLM Gateway",
+  modelsDir: "providers/llmgateway-providers/models",
+  async fetchModels() {
+    return fetchLLMGatewayModels(`${API_ENDPOINT}?mapped=true`);
+  },
+  parseModels(raw) {
+    const data = LLMGatewayResponse.parse(raw).data;
+    // A deployment without the mapped view ignores the query param and returns
+    // aggregated root IDs (no provider prefix); syncing those here would wipe
+    // the provider-pinned catalog, so refuse to proceed. An empty response (or
+    // one left empty after filtering) would silently do the same via the
+    // delete-missing pass, so it is equally fatal.
+    if (data.length === 0 || !data.every((model) => model.id.includes("/"))) {
+      throw new Error("LLM Gateway mapped view unavailable: response is empty or contains unprefixed model ids");
+    }
+    // llmgateway/custom is the BYO-model placeholder and llmgateway/auto the
+    // auto-router; pinning either to a provider is meaningless in this catalog
+    // (the aggregated llmgateway provider carries `auto`).
+    const mapped = data.filter((model) => !model.id.startsWith("llmgateway/") && textOnly(model));
+    if (mapped.length === 0) {
+      throw new Error("LLM Gateway mapped view returned no text models");
+    }
+    // Every mapped entry is one specific provider deployment whose single
+    // providers[] mapping drives capabilities and reasoning controls. A kept
+    // entry with zero or several mappings would make the builder silently fall
+    // back to noisy supported_parameters / sibling defaults, so fail loudly.
+    const malformed = mapped.filter((model) => model.providers?.length !== 1);
+    if (malformed.length > 0) {
+      throw new Error(
+        `LLM Gateway mapped view returned entries without exactly one provider mapping: ${
+          malformed.map((model) => model.id).join(", ")
+        }`,
+      );
+    }
+    return mapped;
+  },
+  translateModel(model, context) {
+    const translated = buildLLMGatewayMappedModel(model, context.existing(model.id));
+    if (translated === undefined) {
+      return undefined;
+    }
+    return { id: model.id, model: translated, header: toggleHeader(translated) };
+  },
+  sourceID(model) {
+    return model.id;
   },
 } satisfies SyncProvider<LLMGatewayModel>;
 
@@ -106,12 +225,75 @@ function modalities(values: string[], fallback: Modality[]): Modality[] {
   return [...new Set(result.length > 0 ? result : fallback)];
 }
 
-function resolveLLMGatewayBaseModel(model: LLMGatewayModel) {
-  const alias = BASE_MODEL_ALIASES[model.id];
+// Modalities as served by a specific deployment: a mapping without vision must
+// not carry image/pdf input, regardless of what the model-level architecture
+// claims — attachment=false with image input is contradictory.
+function deploymentModalities(model: LLMGatewayModel, vision: boolean | undefined) {
+  const base = defaultModalities(model);
+  if (vision !== false) {
+    return base;
+  }
+  const input = base.input.filter((value) => value !== "image" && value !== "pdf");
+  return {
+    input: input.length > 0 ? input : (["text"] satisfies Modality[]),
+    output: base.output,
+  };
+}
+
+const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
+const AGGREGATED_MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "providers", "llmgateway", "models");
+const canonicalOutputLimitByID = new Map<string, number | undefined>();
+
+interface SiblingCuration {
+  reasoning_options?: SyncedFullModel["reasoning_options"];
+  interleaved?: SyncedFullModel["interleaved"];
+  cost_tiers?: NonNullable<SyncedFullModel["cost"]>["tiers"];
+}
+
+const siblingCurationByID = new Map<string, SiblingCuration>();
+
+// The aggregated llmgateway catalog curates reasoning controls, the reasoning
+// side-channel, and context pricing tiers for the same gateway surface; mapped
+// deployments of the same root model reuse them when the deployment does not
+// declare its own.
+function siblingCuration(rootID: string): SiblingCuration {
+  let curation = siblingCurationByID.get(rootID);
+  if (curation === undefined) {
+    const filePath = path.join(AGGREGATED_MODELS_DIR, `${rootID}.toml`);
+    const authored = existsSync(filePath)
+      ? Bun.TOML.parse(readFileSync(filePath, "utf8")) as SiblingCuration & {
+          cost?: { tiers?: NonNullable<SyncedFullModel["cost"]>["tiers"] };
+        }
+      : undefined;
+    curation = {
+      reasoning_options: authored?.reasoning_options?.length ? authored.reasoning_options : undefined,
+      interleaved: authored?.interleaved,
+      cost_tiers: authored?.cost?.tiers,
+    };
+    siblingCurationByID.set(rootID, curation);
+  }
+  return curation;
+}
+
+// Whether the canonical metadata declares limit.output; factored entries can
+// only omit their own output override when the base has one to inherit.
+function canonicalOutputLimit(modelID: string) {
+  if (!canonicalOutputLimitByID.has(modelID)) {
+    const filePath = path.join(MODELS_DIR, `${modelID}.toml`);
+    const metadata = existsSync(filePath)
+      ? Bun.TOML.parse(readFileSync(filePath, "utf8")) as { limit?: { output?: number } }
+      : undefined;
+    canonicalOutputLimitByID.set(modelID, metadata?.limit?.output);
+  }
+  return canonicalOutputLimitByID.get(modelID);
+}
+
+function resolveLLMGatewayBaseModel(model: LLMGatewayModel, modelID = model.id) {
+  const alias = BASE_MODEL_ALIASES[modelID];
   if (alias !== undefined) return alias;
   if (model.family === undefined) return undefined;
   const prefix = CANONICAL_FAMILY_ALIASES[model.family] ?? model.family;
-  return resolveCanonicalBaseModel(`${prefix}/${model.id}`);
+  return resolveModelMetadataBaseModel(`${prefix}/${modelID}`);
 }
 
 function inferFamily(model: LLMGatewayModel, name: string) {
@@ -133,17 +315,23 @@ function inferFamily(model: LLMGatewayModel, name: string) {
 export function buildLLMGatewayModel(
   model: LLMGatewayModel,
   existing: ExistingModel | undefined,
-): SyncedModel {
+): SyncedModel | undefined {
   const prompt = price(model.pricing.prompt);
   const completion = price(model.pricing.completion);
   const reasoning = model.supported_parameters.includes("reasoning")
     || model.supported_parameters.includes("include_reasoning");
-  const context = model.context_length > 0
-    ? model.context_length
-    : existing?.limit?.context ?? model.context_length;
+  const reasoningOptions = llmGatewayReasoningOptions(model, existing);
+  const reported = model.context_length ?? 0;
+  // A missing/zero context must never be authored as limit.context = 0:
+  // factored entries leave it unset and inherit the base, and unfactored
+  // creates are skipped entirely. An authored 0 on the existing file is
+  // equally unusable and must not be re-stamped.
+  const servedContext = reported > 0 ? reported : undefined;
+  const context = servedContext ?? (existing?.limit?.context || undefined);
 
-  // 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
@@ -158,15 +346,24 @@ export function buildLLMGatewayModel(
         tiers: existing?.cost?.tiers,
       }
     : existing?.cost;
-  const limit = {
-    context,
-    input: existing?.limit?.input,
-    output: existing?.limit?.output ?? context,
-  };
+  // Authored limits carry only known-positive values — never the zero/absent
+  // `reported` fallback.
+  const limit = context !== undefined
+    ? {
+        context,
+        input: existing?.limit?.input,
+        output: (existing?.limit?.output || undefined) ?? context,
+      }
+    : undefined;
 
   // Existing factored model: refresh cost + limit, keep every authored override
   // as-is (undefined fields keep inheriting the base model).
   if (existing?.base_model !== undefined) {
+    const factoredLimit = {
+      context,
+      input: existing.limit?.input,
+      output: existing.limit?.output ?? context,
+    };
     return factorBaseModel(
       existing.base_model,
       {
@@ -179,10 +376,11 @@ export function buildLLMGatewayModel(
           tool_call: existing.tool_call,
           structured_output: existing.structured_output,
           open_weights: existing.open_weights,
-          limit,
+          limit: factoredLimit,
           modalities: existing.modalities,
         }),
         reasoning: existing.reasoning,
+        reasoning_options: reasoningOptions,
         temperature: existing.temperature,
         tool_call: existing.tool_call,
         structured_output: existing.structured_output,
@@ -190,16 +388,22 @@ export function buildLLMGatewayModel(
         interleaved: existing.interleaved,
         knowledge: existing.knowledge,
         modalities: existing.modalities,
-        limit,
+        limit: factoredLimit,
         cost,
       },
-      limit,
+      factoredLimit,
       existing.base_model_omit,
     );
   }
 
   // Existing full model: refresh cost + limit, preserve curated metadata.
   if (existing !== undefined) {
+    // With no usable context from the API or the file there is nothing valid
+    // to author, and skipping would hand the file to the delete-missing pass —
+    // fail loudly rather than write limit.context = 0.
+    if (limit === undefined) {
+      throw new Error(`LLM Gateway entry ${model.id} has no usable context to author`);
+    }
     return {
       name: existing.name ?? model.name,
       description: existing.description ?? describeModel({
@@ -218,6 +422,7 @@ 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,
@@ -240,11 +445,20 @@ 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
-  // modality data are unreliable here and should be hand-reviewed.
+  // modality data are unreliable here and should be hand-reviewed. Without a
+  // positive served context there is nothing usable to author, so skip.
+  if (servedContext === undefined) {
+    return undefined;
+  }
+  const createdLimit = limit ?? { context: servedContext, input: undefined, output: servedContext };
   const { input, output } = defaultModalities(model);
   return {
     name: model.name,
@@ -257,7 +471,7 @@ export function buildLLMGatewayModel(
         || model.supported_parameters.includes("tool_choice"),
       structured_output: model.structured_outputs ?? false,
       open_weights: false,
-      limit,
+      limit: createdLimit,
       modalities: { input, output },
     }),
     family: inferFamily(model, model.name),
@@ -265,17 +479,293 @@ 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"),
     structured_output: model.structured_outputs ?? false,
     open_weights: false,
     cost,
-    limit,
+    limit: createdLimit,
     modalities: { input, output },
   } satisfies SyncedFullModel;
 }
 
+export function buildLLMGatewayMappedModel(
+  model: LLMGatewayModel,
+  existing: ExistingModel | undefined,
+): SyncedModel | undefined {
+  // Mapped entries carry exactly one provider mapping; its capability flags
+  // describe that specific deployment, unlike the aggregated view where
+  // supported_parameters are too noisy to trust.
+  const mapping = model.providers?.[0];
+  const rootID = model.id.split("/").slice(1).join("/");
+  const prompt = price(model.pricing.prompt);
+  const completion = price(model.pricing.completion);
+  // The mapping's flag stays authoritative on resyncs too, so the written
+  // reasoning boolean and the reasoning_options derived from it always move
+  // together; prior curation only fills in when the mapping is silent, then
+  // the noisy supported_parameters signal as a last resort.
+  const reasoning = mapping?.reasoning
+    ?? existing?.reasoning
+    ?? (model.supported_parameters.includes("reasoning")
+      || model.supported_parameters.includes("include_reasoning"));
+  // The exact reasoning_effort values this deployment accepts. A deployment
+  // whose only accepted effort is "none" exposes a plain on/off switch (the
+  // gateway honours it through the thinking toggle), not effort tiers.
+  const deploymentOptions = mapping?.reasoning_efforts?.length
+    ? mapping.reasoning_efforts.length === 1 && mapping.reasoning_efforts[0] === "none"
+      ? [{ type: "toggle" as const }]
+      : [{ type: "effort" as const, values: mapping.reasoning_efforts }]
+    : undefined;
+  // Deployment-declared efforts own the effort/toggle surface; curation falls
+  // back from non-empty options on this file to the aggregated llmgateway
+  // catalog's controls for the same root model on the same gateway surface.
+  // Curated non-effort controls (e.g. budget_tokens for $.reasoning.max_tokens,
+  // which this host serves regardless of the effort list) survive alongside
+  // deployment efforts instead of being wiped by them. A curated [] counts as
+  // unknown so a bad first stamp is not sticky. Non-reasoning deployments
+  // carry none; the same applies to the interleaved reasoning side-channel.
+  const sibling = siblingCuration(rootID);
+  const curatedOptions = (existing?.reasoning_options?.length ? existing.reasoning_options : undefined)
+    ?? sibling.reasoning_options;
+  const reasoningOptions = reasoning
+    ? deploymentOptions !== undefined
+      ? [
+          ...(curatedOptions ?? []).filter((option) => option.type !== "effort" && option.type !== "toggle"),
+          ...deploymentOptions,
+        ]
+      : curatedOptions
+    : undefined;
+  const interleaved = reasoning
+    ? existing?.interleaved ?? sibling.interleaved
+    : undefined;
+  const reported = model.context_length ?? 0;
+  // Same zero-context rule as the aggregated builder: never author 0, inherit
+  // on factored entries, skip unfactored creates. An authored 0 on the
+  // existing file is equally unusable.
+  const servedContext = reported > 0 ? reported : undefined;
+  const context = servedContext ?? (existing?.limit?.context || undefined);
+
+  const cost = prompt !== undefined && completion !== undefined
+    ? {
+        input: prompt,
+        output: completion,
+        reasoning: reasoning ? nonZeroPrice(model.pricing.internal_reasoning) ?? existing?.cost?.reasoning : existing?.cost?.reasoning,
+        cache_read: nonZeroPrice(model.pricing.input_cache_read) ?? existing?.cost?.cache_read,
+        cache_write: nonZeroPrice(model.pricing.input_cache_write) ?? existing?.cost?.cache_write,
+        // The gateway API does not expose context pricing tiers, so authored
+        // tiers stick and new files seed from the aggregated sibling's curated
+        // tiers rather than silently under-stating long-context pricing.
+        tiers: existing?.cost?.tiers ?? sibling.cost_tiers,
+      }
+    : existing?.cost;
+  // The gateway's max_output is the deployment's real served limit, so it wins
+  // over inherited/authored values, unlike the aggregated view.
+  const servedOutput = (model.max_output || undefined) ?? (existing?.limit?.output || undefined);
+  // Authored limits carry only known-positive values — never the zero/absent
+  // `reported` fallback.
+  const limit = context !== undefined
+    ? {
+        context,
+        input: existing?.limit?.input,
+        output: servedOutput ?? context,
+      }
+    : undefined;
+
+  // Existing factored model: refresh cost + limit, keep every authored override
+  // as-is. Unlike the aggregated provider, the name override must be carried
+  // forward: mapped names disambiguate deployments of the same model (e.g.
+  // "GPT-5.5 (Azure)" vs "GPT-5.5 (OpenAI)") and must not collapse back to the
+  // base metadata name.
+  if (existing?.base_model !== undefined) {
+    // Mirror the brand-new factored path: without a served or authored output,
+    // keep inheriting the base's output rather than stamping context over it.
+    const factoredLimit = {
+      context,
+      input: existing.limit?.input,
+      output: servedOutput ?? (canonicalOutputLimit(existing.base_model) !== undefined ? undefined : context),
+    };
+    // Deployment capability flags keep their create-path authority on
+    // resyncs: a mapping that gains or loses reasoning/vision/tools/structured
+    // outputs realigns the written flags together with the reasoning_options
+    // computed from them, instead of freezing stale curation forever.
+    return factorBaseModel(
+      existing.base_model,
+      {
+        name: existing.name ?? model.name,
+        attachment: mapping?.vision ?? existing.attachment,
+        // No describeModel fallback: synthesizing a description here would
+        // stamp a sticky generic override on every name-pinned factored entry;
+        // leaving it unset keeps inheriting the lab text from the base.
+        description: existing.description,
+        reasoning: mapping?.reasoning ?? existing.reasoning,
+        reasoning_options: reasoningOptions,
+        temperature: existing.temperature,
+        tool_call: mapping?.tools ?? existing.tool_call,
+        structured_output: model.structured_outputs ?? existing.structured_output,
+        status: existing.status,
+        interleaved,
+        knowledge: existing.knowledge,
+        // Vision realigns modalities in both directions: false strips
+        // image/pdf, true clears any stale stripped override so the base's
+        // richer inputs inherit again; only a silent mapping keeps curation.
+        modalities: mapping?.vision === undefined
+          ? existing.modalities
+          : mapping.vision
+            ? undefined
+            : deploymentModalities(model, false),
+        limit: factoredLimit,
+        cost,
+      },
+      factoredLimit,
+      existing.base_model_omit,
+    );
+  }
+
+  // Existing full model: refresh cost + limit, preserve curated metadata.
+  // Capability flags follow the same rule as the factored path above: the
+  // deployment mapping wins, curation fills the gaps.
+  if (existing !== undefined) {
+    // With no usable context from the API or the file there is nothing valid
+    // to author, and skipping would hand the file to the delete-missing pass —
+    // fail loudly rather than write limit.context = 0.
+    if (limit === undefined) {
+      throw new Error(`LLM Gateway mapped entry ${model.id} has no usable context to author`);
+    }
+    const resolved = {
+      attachment: mapping?.vision ?? existing.attachment ?? false,
+      tool_call: mapping?.tools ?? existing.tool_call ?? false,
+      structured_output: model.structured_outputs ?? existing.structured_output,
+      // Same bidirectional vision rule as the factored path; with no base to
+      // inherit from, a declared vision recomputes from the served
+      // architecture instead of clearing.
+      modalities: mapping?.vision === undefined
+        ? existing.modalities ?? deploymentModalities(model, undefined)
+        : deploymentModalities(model, mapping.vision),
+    };
+    return {
+      name: existing.name ?? model.name,
+      description: existing.description ?? describeModel({
+        id: model.id,
+        name: existing.name ?? model.name,
+        family: existing.family,
+        reasoning,
+        tool_call: resolved.tool_call,
+        structured_output: resolved.structured_output,
+        open_weights: existing.open_weights,
+        limit,
+        modalities: resolved.modalities,
+      }),
+      family: existing.family,
+      release_date: existing.release_date ?? dateFromTimestamp(model.created),
+      last_updated: existing.last_updated ?? dateFromTimestamp(model.created),
+      attachment: resolved.attachment,
+      reasoning,
+      reasoning_options: reasoningOptions,
+      temperature: existing.temperature ?? false,
+      tool_call: resolved.tool_call,
+      structured_output: resolved.structured_output,
+      knowledge: existing.knowledge,
+      open_weights: existing.open_weights ?? false,
+      status: existing.status,
+      interleaved,
+      cost,
+      limit,
+      modalities: resolved.modalities,
+    } satisfies SyncedFullModel;
+  }
+
+  // Brand-new model with a reviewed metadata entry: factor against the
+  // canonical base. The mapped ID is `serving-provider/model-id` and the
+  // serving provider is unrelated to the originating lab, so resolve the base
+  // from the root model ID + family, and keep the disambiguating name. The
+  // mapping's own capability flags describe this specific deployment, so they
+  // go in as overrides (factorBaseModel drops the ones equal to the base).
+  const canonical = resolveLLMGatewayBaseModel(model, rootID);
+  if (canonical !== undefined) {
+    const factoredLimit = {
+      context,
+      input: undefined,
+      // Without a served limit, inherit the base's output; only fall back to
+      // context when the base declares none (output is required downstream).
+      output: model.max_output ?? (canonicalOutputLimit(canonical) !== undefined ? undefined : context),
+    };
+    return factorBaseModel(canonical, {
+      name: model.name,
+      attachment: mapping?.vision,
+      reasoning: mapping?.reasoning,
+      reasoning_options: reasoningOptions,
+      interleaved,
+      tool_call: mapping?.tools,
+      structured_output: model.structured_outputs,
+      // A deployment without vision must not inherit image/pdf inputs from
+      // the base — attachment=false with image input is contradictory.
+      modalities: mapping?.vision === false ? deploymentModalities(model, false) : undefined,
+      limit: factoredLimit,
+      cost,
+    }, factoredLimit);
+  }
+
+  // Brand-new model without metadata: best-effort translation. The mapping's
+  // own capability flags are reliable here; modalities mirror the mapping too.
+  // Without a positive served context there is nothing usable to author.
+  if (servedContext === undefined) {
+    return undefined;
+  }
+  const createdLimit = limit ?? { context: servedContext, input: undefined, output: servedOutput ?? servedContext };
+  const { input, output } = deploymentModalities(model, mapping?.vision);
+  return {
+    name: model.name,
+    description: describeModel({
+      id: model.id,
+      name: model.name,
+      family: inferFamily(model, model.name),
+      reasoning,
+      tool_call: mapping?.tools ?? false,
+      structured_output: model.structured_outputs ?? false,
+      open_weights: false,
+      limit: createdLimit,
+      modalities: { input, output },
+    }),
+    family: inferFamily(model, model.name),
+    release_date: dateFromTimestamp(model.created),
+    last_updated: dateFromTimestamp(model.created),
+    attachment: mapping?.vision ?? input.some((value) => value !== "text"),
+    reasoning,
+    reasoning_options: reasoningOptions,
+    interleaved,
+    temperature: model.supported_parameters.includes("temperature"),
+    tool_call: mapping?.tools ?? false,
+    structured_output: model.structured_outputs ?? false,
+    open_weights: false,
+    cost,
+    limit: createdLimit,
+    modalities: { input, output },
+  } 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

@@ -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;

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

@@ -55,8 +55,11 @@ export const ofox = {
   name: "Ofox",
   modelsDir: "providers/ofox/models",
   skipCreates: true,
-  trackMissingModels: false,
+  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.`,