[github-models/deepseek] Add reasoning options

fix: remove GLM-5.2 from standard Z.AI APIs

.github/workflows/close-stale-pull-requests.yml

@@ -0,0 +1,68 @@
+name: Close stale pull requests
+
+on:
+  schedule:
+    - cron: "17 3 * * *"
+  workflow_dispatch:
+
+permissions:
+  issues: write
+  pull-requests: write
+
+jobs:
+  close-stale-pull-requests:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/github-script@v8
+        env:
+          REVIEWER: rekram1-node
+        with:
+          script: |
+            const { owner, repo } = context.repo
+            const now = Date.now()
+            const weekAgo = now - 7 * 24 * 60 * 60 * 1000
+            const monthAgo = now - 30 * 24 * 60 * 60 * 1000
+
+            const pulls = await github.paginate(github.rest.pulls.list, {
+              owner,
+              repo,
+              state: "open",
+              per_page: 100,
+            })
+
+            const feedbackPulls = new Set()
+            for (const qualifier of ["commenter", "reviewed-by"]) {
+              const results = await github.paginate(
+                github.rest.search.issuesAndPullRequests,
+                {
+                  q: `repo:${owner}/${repo} is:pr is:open ${qualifier}:${process.env.REVIEWER}`,
+                  per_page: 100,
+                },
+              )
+
+              for (const result of results) feedbackPulls.add(result.number)
+            }
+
+            for (const pull of pulls) {
+              const updatedAt = Date.parse(pull.updated_at)
+              const monthStale = updatedAt < monthAgo
+              const feedbackStale = updatedAt < weekAgo && feedbackPulls.has(pull.number)
+              if (!monthStale && !feedbackStale) continue
+
+              const reason = monthStale
+                ? "it has not been updated in 30 days"
+                : `it has not been updated in 7 days after feedback from @${process.env.REVIEWER}`
+
+              await github.rest.issues.createComment({
+                owner,
+                repo,
+                issue_number: pull.number,
+                body: `Closing this pull request as stale because ${reason}. Feel free to reopen it or submit a new pull request if the work is resumed.`,
+              })
+              await github.rest.pulls.update({
+                owner,
+                repo,
+                pull_number: pull.number,
+                state: "closed",
+              })
+            }

.github/workflows/sync-models.yml

@@ -63,7 +63,9 @@ jobs:
       - name: Sync model catalogs
         run: bun models:sync ${{ matrix.provider }}
         env:
+          BASETEN_API_KEY: ${{ secrets.BASETEN_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
+          VENICE_API_KEY: ${{ secrets.VENICE_API_KEY }}
           GOOGLE_API_KEY: ${{ secrets.GOOGLE_API_KEY }}
           GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
           GOOGLE_GENERATIVE_AI_API_KEY: ${{ secrets.GOOGLE_GENERATIVE_AI_API_KEY }}
@@ -81,7 +83,7 @@ jobs:
           LABELS: automation,model-sync,provider:${{ matrix.provider }}
           TITLE: "chore(sync): update ${{ matrix.name }} model catalog"
         run: |
-          if [ -z "$(git status --porcelain -- providers)" ]; then
+          if [ -z "$(git status --porcelain -- models providers)" ]; then
             echo "No model catalog changes found."
             exit 0
           fi
@@ -90,7 +92,7 @@ jobs:
           git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
           git fetch --no-tags --depth=1 origin "+refs/heads/$BRANCH:refs/remotes/origin/$BRANCH" || true
           git checkout -B "$BRANCH"
-          git add providers
+          git add models providers
           git commit -m "$TITLE"
           git push --force-with-lease origin "$BRANCH"
 

models/alibaba/qwen3.7-plus.toml

@@ -10,9 +10,9 @@ knowledge = "2025-04"
 open_weights = false
 
 [limit]
-context = 131_072
-output = 16_384
+context = 1_000_000
+output = 64_000
 
 [modalities]
-input = ["text"]
+input = ["text", "image"]
 output = ["text"]

models/anthropic/claude-fable-5.toml

@@ -0,0 +1,18 @@
+name = "Claude Fable 5"
+family = "claude-fable"
+release_date = "2026-06-09"
+last_updated = "2026-06-09"
+attachment = true
+reasoning = true
+temperature = false
+tool_call = true
+open_weights = false
+knowledge = "2026-01-31"
+
+[limit]
+context = 1_000_000
+output = 128_000
+
+[modalities]
+input = ["text", "image", "pdf"]
+output = ["text"]

models/cohere/command-a-plus-05-2026.toml

@@ -0,0 +1,19 @@
+name = "Command A Plus"
+family = "command-a"
+release_date = "2026-05-20"
+last_updated = "2026-06-09"
+attachment = true
+reasoning = true
+temperature = true
+knowledge = "2025-04-01"
+tool_call = true
+open_weights = true
+structured_output = true
+
+[limit]
+context = 128_000
+output = 64_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

models/cohere/north-mini-code-1-0.toml

@@ -0,0 +1,19 @@
+name = "North Mini Code"
+family = "north"
+release_date = "2026-06-09"
+last_updated = "2026-06-09"
+attachment = false
+reasoning = true
+temperature = true
+structured_output = true
+knowledge = "2025-09-23"
+tool_call = true
+open_weights = true
+
+[limit]
+context = 256_000
+output = 64_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/google/gemini-2.5-flash.toml

@@ -1,7 +1,7 @@
 name = "Gemini 2.5 Flash"
 family = "gemini-flash"
-release_date = "2025-03-20"
-last_updated = "2025-06-05"
+release_date = "2025-06-17"
+last_updated = "2025-06-17"
 attachment = true
 reasoning = true
 temperature = true

models/google/gemini-2.5-pro-tts.toml

@@ -0,0 +1,18 @@
+name = "Gemini 2.5 Pro TTS"
+family = "gemini-pro"
+release_date = "2025-09-30"
+last_updated = "2025-12-10"
+attachment = false
+reasoning = false
+temperature = false
+tool_call = false
+knowledge = "2025-01"
+open_weights = false
+
+[limit]
+context = 32_768
+output = 16_384
+
+[modalities]
+input = ["text"]
+output = ["audio"]

models/google/gemini-2.5-pro.toml

@@ -1,7 +1,7 @@
 name = "Gemini 2.5 Pro"
 family = "gemini-pro"
-release_date = "2025-03-20"
-last_updated = "2025-06-05"
+release_date = "2025-06-17"
+last_updated = "2025-06-17"
 attachment = true
 reasoning = true
 temperature = true

models/mistral/mistral-large-2411.toml

@@ -1,7 +1,7 @@
 name = "Mistral Large 2.1"
 family = "mistral-large"
-release_date = "2024-11-01"
-last_updated = "2024-11-04"
+release_date = "2024-11-18"
+last_updated = "2024-11-18"
 attachment = false
 reasoning = false
 temperature = true

models/moonshotai/kimi-k2.5.toml

@@ -1,5 +1,5 @@
 name = "Kimi K2.5"
-family = "kimi-k2.5"
+family = "kimi-k2"
 release_date = "2026-01"
 last_updated = "2026-01"
 attachment = false

models/moonshotai/kimi-k2.6.toml

@@ -1,5 +1,5 @@
 name = "Kimi K2.6"
-family = "kimi-k2.6"
+family = "kimi-k2"
 release_date = "2026-04-21"
 last_updated = "2026-04-21"
 attachment = true

models/moonshotai/kimi-k2.7-code.toml

@@ -0,0 +1,23 @@
+name = "Kimi K2.7 Code"
+family = "kimi-k2"
+release_date = "2026-06-12"
+last_updated = "2026-06-12"
+attachment = true
+reasoning = true
+temperature = false
+tool_call = true
+structured_output = true
+knowledge = "2025-01"
+open_weights = true
+
+[limit]
+context = 262_144
+output = 262_144
+
+[modalities]
+input = ["text", "image", "video"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/moonshotai/Kimi-K2.7-Code"

models/openai/gpt-5.5-instant.toml

@@ -0,0 +1,19 @@
+name = "GPT-5.5 Instant"
+release_date = "2026-05-05"
+last_updated = "2026-05-28"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+structured_output = true
+knowledge = "2025-12-01"
+open_weights = false
+
+[limit]
+context = 400_000
+input = 400_000
+output = 128_000
+
+[modalities]
+input = ["text", "image", "pdf"]
+output = ["text"]

models/sarvam/sarvam-105b.toml

@@ -0,0 +1,17 @@
+name = "Sarvam 105B"
+family = "sarvam"
+release_date = "2025-09-01"
+last_updated = "2025-09-01"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 131_072
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/sarvam/sarvam-30b.toml

@@ -0,0 +1,17 @@
+name = "Sarvam 30B"
+family = "sarvam"
+release_date = "2026-02-18"
+last_updated = "2026-02-18"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+open_weights = true
+
+[limit]
+context = 128_000
+output = 128_000
+
+[modalities]
+input = ["text"]
+output = ["text"]

models/stepfun/step-3.7-flash.toml

@@ -0,0 +1,22 @@
+name = "Step 3.7 Flash"
+release_date = "2026-05-29"
+last_updated = "2026-05-29"
+attachment = true
+reasoning = true
+temperature = true
+tool_call = true
+knowledge = "2026-01-01"
+open_weights = true
+
+[limit]
+context = 256_000
+input = 256_000
+output = 256_000
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/stepfun-ai/Step-3.7-Flash"

models/xai/grok-4.20-0309-non-reasoning.toml

@@ -6,10 +6,11 @@ attachment = true
 reasoning = false
 temperature = true
 tool_call = true
+structured_output = true
 open_weights = false
 
 [limit]
-context = 2_000_000
+context = 1_000_000
 output = 30_000
 
 [modalities]

models/xai/grok-4.20-0309-reasoning.toml

@@ -6,10 +6,11 @@ attachment = true
 reasoning = true
 temperature = true
 tool_call = true
+structured_output = true
 open_weights = false
 
 [limit]
-context = 2_000_000
+context = 1_000_000
 output = 30_000
 
 [modalities]

models/xai/grok-4.3.toml

@@ -6,6 +6,7 @@ attachment = true
 reasoning = true
 temperature = true
 tool_call = true
+structured_output = true
 open_weights = false
 
 [limit]

models/xiaomi/mimo-v2.5-pro-ultraspeed.toml

@@ -0,0 +1,22 @@
+name = "MiMo-V2.5-Pro-UltraSpeed"
+family = "mimo"
+release_date = "2026-06-08"
+last_updated = "2026-06-09"
+attachment = false
+reasoning = true
+temperature = true
+tool_call = true
+knowledge = "2024-12"
+open_weights = true
+
+[limit]
+context = 1_048_576
+output = 131_072
+
+[modalities]
+input = ["text"]
+output = ["text"]
+
+[[weights]]
+label = "Hugging Face"
+url = "https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash"

models/zhipuai/glm-5.2.toml

@@ -0,0 +1,18 @@
+name = "GLM-5.2"
+family = "glm"
+release_date = "2026-06-13"
+last_updated = "2026-06-13"
+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

@@ -18,11 +18,12 @@
     "test": "bun test",
     "validate": "bun ./packages/core/script/validate.ts",
     "compare:migrations": "bun ./packages/core/script/compare-model-migrations.ts",
+    "baseten:sync": "bun ./packages/core/script/sync-models.ts baseten",
     "cloudflare:sync": "bun ./packages/core/script/sync-models.ts cloudflare-workers-ai",
     "chutes:generate": "bun ./packages/core/script/generate-chutes.ts",
     "databricks:generate": "bun ./packages/core/script/generate-databricks.ts",
     "helicone:generate": "bun ./packages/core/script/generate-helicone.ts",
-    "venice:generate": "bun ./packages/core/script/generate-venice.ts",
+    "venice:sync": "bun ./packages/core/script/sync-models.ts venice",
     "vercel:generate": "bun ./packages/core/script/sync-models.ts vercel",
     "wandb:generate": "bun ./packages/core/script/generate-wandb.ts",
     "digitalocean:generate": "bun ./packages/core/script/generate-digitalocean.ts",

packages/core/script/generate-chutes.ts

@@ -13,7 +13,7 @@ import { z } from "zod";
 import path from "node:path";
 import { existsSync, readFileSync } from "node:fs";
 import { mkdir } from "node:fs/promises";
-import { ModelFamilyValues } from "../src/family.js";
+import { inferKimiFamily, ModelFamilyValues } from "../src/family.js";
 
 const API_ENDPOINT = "https://llm.chutes.ai/v1/models";
 const MODEL_METADATA_DIR = path.join(import.meta.dirname, "..", "..", "..", "models");
@@ -261,6 +261,9 @@ function matchesFamily(target: string, family: string): boolean {
 }
 
 function inferFamily(modelId: string, modelName: string): string | undefined {
+  const kimiFamily = inferKimiFamily(modelId, modelName);
+  if (kimiFamily !== undefined) return kimiFamily;
+
   const sortedFamilies = [...ModelFamilyValues].sort((a, b) => b.length - a.length);
 
   // First pass: try exact substring matches

packages/core/script/generate-digitalocean.ts

@@ -26,7 +26,7 @@
 import { z } from "zod";
 import path from "node:path";
 import { mkdir } from "node:fs/promises";
-import { ModelFamilyValues } from "../src/family.js";
+import { inferKimiFamily, ModelFamilyValues } from "../src/family.js";
 
 const MODELS_API = "https://api.digitalocean.com/v2/gen-ai/models";
 const PRICING_API = "https://www.digitalocean.com/api/static-content/v1/products";
@@ -142,7 +142,7 @@ const PRICING_NAME_MAP: Record<string, string> = {
   // DO-hosted
   "qwen3-32b": "alibaba-qwen3-32b",
   "minimax m2.5 (public preview)": "minimax-m2.5",
-  "kimi k2.5": "kimi-k2.5",
+  "kimi k2.5": "kimi-k2",
   "nvidia nemotron 3 super 120b (public preview)": "nvidia-nemotron-3-super-120b",
   "glm 5": "glm-5",
 };
@@ -311,6 +311,9 @@ function formatNumber(n: number): string {
 }
 
 function inferFamily(modelId: string, modelName: string): string | undefined {
+  const kimiFamily = inferKimiFamily(modelId, modelName);
+  if (kimiFamily !== undefined) return kimiFamily;
+
   const sorted = [...ModelFamilyValues].sort((a, b) => b.length - a.length);
   const targets = [modelId.toLowerCase(), modelName.toLowerCase()];
   for (const family of sorted) {

packages/core/script/generate-friendli.ts

@@ -4,6 +4,8 @@ import { mkdir } from "node:fs/promises";
 import path from "node:path";
 import { z } from "zod";
 
+import { inferKimiFamily } from "../src/family.js";
+
 // Friendli API endpoint
 const API_ENDPOINT = "https://api.friendli.ai/serverless/v1/models";
 
@@ -53,6 +55,9 @@ const familyPatterns: [RegExp, string][] = [
 ];
 
 function inferFamily(modelId: string, modelName: string): string | undefined {
+  const kimiFamily = inferKimiFamily(modelId, modelName);
+  if (kimiFamily !== undefined) return kimiFamily;
+
   for (const [pattern, family] of familyPatterns) {
     if (pattern.test(modelId) || pattern.test(modelName)) {
       return family;

packages/core/script/generate-venice.ts

@@ -1,653 +0,0 @@
-#!/usr/bin/env bun
-
-import { z } from "zod";
-import path from "node:path";
-import { readdir } from "node:fs/promises";
-import { ModelFamilyValues } from "../src/family.js";
-
-// Venice API endpoint
-const API_ENDPOINT = "https://api.venice.ai/api/v1/models?type=text";
-
-// Zod schemas for API response validation
-const Capabilities = z
-  .object({
-    optimizedForCode: z.boolean().optional(),
-    quantization: z.string().optional(),
-    supportsAudioInput: z.boolean().optional(),
-    supportsFunctionCalling: z.boolean().optional(),
-    supportsLogProbs: z.boolean().optional(),
-    supportsReasoning: z.boolean().optional(),
-    supportsResponseSchema: z.boolean().optional(),
-    supportsVideoInput: z.boolean().optional(),
-    supportsVision: z.boolean().optional(),
-    supportsWebSearch: z.boolean().optional(),
-  })
-  .passthrough();
-
-const PricingTier = z.object({ usd: z.number(), diem: z.number().optional() }).passthrough();
-
-const ExtendedPricing = z
-  .object({
-    context_token_threshold: z.number(),
-    input: PricingTier,
-    output: PricingTier,
-    cache_input: PricingTier.optional(),
-    cache_write: PricingTier.optional(),
-  })
-  .passthrough();
-
-const Pricing = z
-  .object({
-    input: PricingTier,
-    output: PricingTier,
-    cache_input: PricingTier.optional(),
-    cache_write: PricingTier.optional(),
-    extended: ExtendedPricing.optional(),
-  })
-  .passthrough();
-
-const ModelSpec = z
-  .object({
-    pricing: Pricing.optional(),
-    availableContextTokens: z.number(),
-    maxCompletionTokens: z.number().optional(),
-    capabilities: Capabilities,
-    constraints: z.any().optional(),
-    name: z.string(),
-    modelSource: z.string().optional(),
-    offline: z.boolean().optional(),
-    privacy: z.string().optional(),
-    traits: z.array(z.string()).optional(),
-  })
-  .passthrough();
-
-const VeniceModel = z
-  .object({
-    created: z.number(),
-    id: z.string(),
-    model_spec: ModelSpec,
-    object: z.string(),
-    owned_by: z.string(),
-    type: z.string(),
-  })
-  .passthrough();
-
-const VeniceResponse = z
-  .object({
-    data: z.array(VeniceModel),
-    object: z.string(),
-    type: z.string(),
-  })
-  .passthrough();
-
-function matchesFamily(target: string, family: string): boolean {
-  const targetLower = target.toLowerCase();
-  const familyLower = family.toLowerCase();
-  let familyIdx = 0;
-
-  for (let i = 0; i < targetLower.length && familyIdx < familyLower.length; i++) {
-    if (targetLower[i] === familyLower[familyIdx]) {
-      familyIdx++;
-    }
-  }
-
-  return familyIdx === familyLower.length;
-}
-
-function inferFamily(modelId: string, modelName: string): string | undefined {
-  const sortedFamilies = [...ModelFamilyValues].sort((a, b) => b.length - a.length);
-
-  for (const family of sortedFamilies) {
-    if (matchesFamily(modelId, family)) {
-      return family;
-    }
-  }
-
-  for (const family of sortedFamilies) {
-    if (matchesFamily(modelName, family)) {
-      return family;
-    }
-  }
-
-  return undefined;
-}
-
-function buildInputModalities(capabilities: z.infer<typeof Capabilities>): string[] {
-  const mods: string[] = ["text"];
-  if (capabilities.supportsVision) mods.push("image");
-  if (capabilities.supportsAudioInput) mods.push("audio");
-  if (capabilities.supportsVideoInput) mods.push("video");
-  return mods;
-}
-
-function formatNumber(n: number): string {
-  if (n >= 1000) {
-    // Format with underscores for readability (e.g., 131_072)
-    return n.toString().replace(/\B(?=(\d{3})+(?!\d))/g, "_");
-  }
-  return n.toString();
-}
-
-function timestampToDate(timestamp: number): string {
-  const date = new Date(timestamp * 1000);
-  return date.toISOString().slice(0, 10);
-}
-
-function getTodayDate(): string {
-  return new Date().toISOString().slice(0, 10);
-}
-
-interface ExistingModel {
-  name?: string;
-  family?: string;
-  attachment?: boolean;
-  reasoning?: boolean;
-  tool_call?: boolean;
-  structured_output?: boolean;
-  temperature?: boolean;
-  knowledge?: string;
-  release_date?: string;
-  last_updated?: string;
-  open_weights?: boolean;
-  interleaved?: boolean | { field: string };
-  status?: string;
-  cost?: {
-    input?: number;
-    output?: number;
-    reasoning?: number;
-    cache_read?: number;
-    cache_write?: number;
-    context_over_200k?: {
-      input?: number;
-      output?: number;
-      cache_read?: number;
-      cache_write?: number;
-      context_min?: number;
-    };
-    tiers?: Array<{
-      tier: {
-        type?: "context";
-        size: number;
-      };
-      input?: number;
-      output?: number;
-      cache_read?: number;
-      cache_write?: number;
-    }>;
-  };
-  limit?: {
-    context?: number;
-    input?: number;
-    output?: number;
-  };
-  modalities?: {
-    input?: string[];
-    output?: string[];
-  };
-  provider?: {
-    npm?: string;
-    api?: string;
-  };
-}
-
-async function loadExistingModel(filePath: string): Promise<ExistingModel | null> {
-  try {
-    const file = Bun.file(filePath);
-    if (!(await file.exists())) {
-      return null;
-    }
-    const toml = await import(filePath, { with: { type: "toml" } }).then(
-      (mod) => mod.default,
-    );
-    return toml as ExistingModel;
-  } catch (e) {
-    console.warn(`Warning: Failed to parse existing file ${filePath}:`, e);
-    return null;
-  }
-}
-
-function getExistingLongContextMin(existing: ExistingModel | null) {
-  return (
-    existing?.cost?.tiers?.find(
-      (tier) =>
-        (tier.tier.type === undefined || tier.tier.type === "context") &&
-        tier.tier.size >= 200_000,
-    )?.tier.size ?? 200_000
-  );
-}
-
-function getExistingLongContextCost(existing: ExistingModel | null) {
-  return (
-    existing?.cost?.tiers?.find(
-      (tier) =>
-        (tier.tier.type === undefined || tier.tier.type === "context") &&
-        tier.tier.size >= 200_000,
-    ) ?? existing?.cost?.context_over_200k
-  );
-}
-
-function getLongContextMin(cost: { context_min?: number }) {
-  return cost.context_min ?? 200_000;
-}
-
-interface MergedModel {
-  name: string;
-  family?: string;
-  attachment: boolean;
-  reasoning: boolean;
-  tool_call: boolean;
-  structured_output?: boolean;
-  temperature: boolean;
-  knowledge?: string;
-  release_date: string;
-  last_updated: string;
-  open_weights: boolean;
-  interleaved?: boolean | { field: string };
-  status?: string;
-  cost?: {
-    input: number;
-    output: number;
-    cache_read?: number;
-    cache_write?: number;
-    context_over_200k?: {
-      input: number;
-      output: number;
-      cache_read?: number;
-      cache_write?: number;
-      context_min?: number;
-    };
-  };
-  limit: {
-    context: number;
-    output: number;
-  };
-  modalities: {
-    input: string[];
-    output: string[];
-  };
-}
-
-function mergeModel(
-  apiModel: z.infer<typeof VeniceModel>,
-  existing: ExistingModel | null,
-): MergedModel {
-  const spec = apiModel.model_spec;
-  const caps = spec.capabilities;
-
-  const contextTokens = spec.availableContextTokens;
-  const outputTokens = spec.maxCompletionTokens ?? Math.floor(contextTokens / 4);
-
-  const openWeights = spec.modelSource?.toLowerCase().includes("huggingface") ?? false;
-
-  const inputModalities = buildInputModalities(caps);
-
-  if (existing?.modalities?.input?.includes("pdf") && !inputModalities.includes("pdf")) {
-    inputModalities.push("pdf");
-  }
-
-  const attachment =
-    caps.supportsVision === true ||
-    caps.supportsAudioInput === true ||
-    caps.supportsVideoInput === true;
-
-  const merged: MergedModel = {
-    name: spec.name,
-    attachment,
-    reasoning: caps.supportsReasoning === true,
-    tool_call: caps.supportsFunctionCalling === true,
-    temperature: true,
-    release_date: timestampToDate(apiModel.created),
-    last_updated: getTodayDate(),
-    open_weights: openWeights,
-    limit: {
-      context: contextTokens,
-      output: outputTokens,
-    },
-    modalities: {
-      input: inputModalities,
-      output: ["text"],
-    },
-  };
-
-  // structured_output only if true
-  if (caps.supportsResponseSchema === true) {
-    merged.structured_output = true;
-  }
-
-  // Cost from API
-  if (spec.pricing) {
-    merged.cost = {
-      input: spec.pricing.input.usd,
-      output: spec.pricing.output.usd,
-      ...(spec.pricing.cache_input && { cache_read: spec.pricing.cache_input.usd }),
-      ...(spec.pricing.cache_write && { cache_write: spec.pricing.cache_write.usd }),
-    };
-
-    // Extended pricing maps to context_over_200k
-    if (spec.pricing.extended) {
-      merged.cost.context_over_200k = {
-        input: spec.pricing.extended.input.usd,
-        output: spec.pricing.extended.output.usd,
-        context_min: spec.pricing.extended.context_token_threshold,
-        ...(spec.pricing.extended.cache_input && { cache_read: spec.pricing.extended.cache_input.usd }),
-        ...(spec.pricing.extended.cache_write && { cache_write: spec.pricing.extended.cache_write.usd }),
-      };
-    }
-  }
-
-  const inferred = inferFamily(apiModel.id, spec.name);
-  merged.family = inferred ?? existing?.family;
-
-  // Preserve manual fields from existing
-  if (existing?.knowledge) {
-    merged.knowledge = existing.knowledge;
-  }
-  if (existing?.interleaved !== undefined) {
-    merged.interleaved = existing.interleaved;
-  }
-  if (existing?.status !== undefined) {
-    merged.status = existing.status;
-  }
-
-  return merged;
-}
-
-function formatToml(model: MergedModel): string {
-  const lines: string[] = [];
-
-  // Basic fields
-  lines.push(`name = "${model.name.replace(/"/g, '\\"')}"`);
-  if (model.family) {
-    lines.push(`family = "${model.family}"`);
-  }
-  lines.push(`attachment = ${model.attachment}`);
-  lines.push(`reasoning = ${model.reasoning}`);
-  lines.push(`tool_call = ${model.tool_call}`);
-  if (model.structured_output !== undefined) {
-    lines.push(`structured_output = ${model.structured_output}`);
-  }
-  lines.push(`temperature = ${model.temperature}`);
-  if (model.knowledge) {
-    lines.push(`knowledge = "${model.knowledge}"`);
-  }
-  lines.push(`release_date = "${model.release_date}"`);
-  lines.push(`last_updated = "${model.last_updated}"`);
-  lines.push(`open_weights = ${model.open_weights}`);
-  if (model.status) {
-    lines.push(`status = "${model.status}"`);
-  }
-
-  // Interleaved section (if present)
-  if (model.interleaved !== undefined) {
-    lines.push("");
-    if (model.interleaved === true) {
-      lines.push(`interleaved = true`);
-    } else if (typeof model.interleaved === "object") {
-      lines.push(`[interleaved]`);
-      lines.push(`field = "${model.interleaved.field}"`);
-    }
-  }
-
-  // Cost section
-  if (model.cost) {
-    lines.push("");
-    lines.push(`[cost]`);
-    lines.push(`input = ${model.cost.input}`);
-    lines.push(`output = ${model.cost.output}`);
-    if (model.cost.cache_read !== undefined) {
-      lines.push(`cache_read = ${model.cost.cache_read}`);
-    }
-    if (model.cost.cache_write !== undefined) {
-      lines.push(`cache_write = ${model.cost.cache_write}`);
-    }
-
-    if (model.cost.context_over_200k) {
-      lines.push("");
-      lines.push(`[[cost.tiers]]`);
-      lines.push(`tier = { size = ${formatNumber(getLongContextMin(model.cost.context_over_200k))} }`);
-      lines.push(`input = ${model.cost.context_over_200k.input}`);
-      lines.push(`output = ${model.cost.context_over_200k.output}`);
-      if (model.cost.context_over_200k.cache_read !== undefined) {
-        lines.push(`cache_read = ${model.cost.context_over_200k.cache_read}`);
-      }
-      if (model.cost.context_over_200k.cache_write !== undefined) {
-        lines.push(`cache_write = ${model.cost.context_over_200k.cache_write}`);
-      }
-    }
-  }
-
-  // Limit section
-  lines.push("");
-  lines.push(`[limit]`);
-  lines.push(`context = ${formatNumber(model.limit.context)}`);
-  lines.push(`output = ${formatNumber(model.limit.output)}`);
-
-  // Modalities section
-  lines.push("");
-  lines.push(`[modalities]`);
-  lines.push(`input = [${model.modalities.input.map((m) => `"${m}"`).join(", ")}]`);
-  lines.push(`output = [${model.modalities.output.map((m) => `"${m}"`).join(", ")}]`);
-
-  return lines.join("\n") + "\n";
-}
-
-interface Changes {
-  field: string;
-  oldValue: string;
-  newValue: string;
-}
-
-function detectChanges(
-  existing: ExistingModel | null,
-  merged: MergedModel,
-): Changes[] {
-  if (!existing) return [];
-
-  const changes: Changes[] = [];
-
-  const compare = (field: string, oldVal: unknown, newVal: unknown) => {
-    const oldStr = JSON.stringify(oldVal);
-    const newStr = JSON.stringify(newVal);
-    if (oldStr !== newStr) {
-      changes.push({
-        field,
-        oldValue: formatValue(oldVal),
-        newValue: formatValue(newVal),
-      });
-    }
-  };
-
-  const formatValue = (val: unknown): string => {
-    if (typeof val === "number") return formatNumber(val);
-    if (Array.isArray(val)) return `[${val.join(", ")}]`;
-    if (val === undefined) return "(none)";
-    return String(val);
-  };
-
-  compare("name", existing.name, merged.name);
-  compare("family", existing.family, merged.family);
-  compare("attachment", existing.attachment, merged.attachment);
-  compare("reasoning", existing.reasoning, merged.reasoning);
-  compare("tool_call", existing.tool_call, merged.tool_call);
-  compare("structured_output", existing.structured_output, merged.structured_output);
-  compare("open_weights", existing.open_weights, merged.open_weights);
-  compare("release_date", existing.release_date, merged.release_date);
-  compare("cost.input", existing.cost?.input, merged.cost?.input);
-  compare("cost.output", existing.cost?.output, merged.cost?.output);
-  compare("cost.cache_read", existing.cost?.cache_read, merged.cost?.cache_read);
-  compare("cost.cache_write", existing.cost?.cache_write, merged.cost?.cache_write);
-  const existingLongContextCost = getExistingLongContextCost(existing);
-  compare("cost.context_over_200k.input", existingLongContextCost?.input, merged.cost?.context_over_200k?.input);
-  compare("cost.context_over_200k.output", existingLongContextCost?.output, merged.cost?.context_over_200k?.output);
-  compare("cost.context_over_200k.cache_read", existingLongContextCost?.cache_read, merged.cost?.context_over_200k?.cache_read);
-  compare("cost.context_over_200k.cache_write", existingLongContextCost?.cache_write, merged.cost?.context_over_200k?.cache_write);
-  compare("limit.context", existing.limit?.context, merged.limit.context);
-  compare("limit.output", existing.limit?.output, merged.limit.output);
-  compare("modalities.input", existing.modalities?.input, merged.modalities.input);
-
-  return changes;
-}
-
-async function main() {
-  const args = process.argv.slice(2);
-  const dryRun = args.includes("--dry-run");
-
-  const modelsDir = path.join(
-    import.meta.dirname,
-    "..",
-    "..",
-    "..",
-    "providers",
-    "venice",
-    "models",
-  );
-
-  // Check for API key from CLI argument or environment variable
-  let apiKey: string | null = null;
-
-  // Check CLI args for --api-key=xxx or --api-key xxx
-  const apiKeyArgIndex = args.findIndex((arg) => arg.startsWith("--api-key"));
-  if (apiKeyArgIndex !== -1) {
-    const arg = args[apiKeyArgIndex];
-    if (arg?.includes("=")) {
-      apiKey = arg.split("=")[1] ?? null;
-    } else if (args[apiKeyArgIndex + 1]) {
-      apiKey = args[apiKeyArgIndex + 1] ?? null;
-    }
-  }
-
-  // Fall back to environment variable
-  if (!apiKey) {
-    apiKey = process.env.VENICE_API_KEY ?? null;
-  }
-
-  const includeAlpha = apiKey !== null;
-
-  if (dryRun) {
-    console.log(
-      `[DRY RUN] Fetching Venice models from API${includeAlpha ? " (including alpha models)" : ""}...`,
-    );
-  } else {
-    console.log(
-      `Fetching Venice models from API${includeAlpha ? " (including alpha models)" : ""}...`,
-    );
-  }
-
-  // Fetch API data
-  const fetchOptions: RequestInit = {};
-  if (apiKey) {
-    fetchOptions.headers = {
-      Authorization: `Bearer ${apiKey}`,
-    };
-  }
-
-  const res = await fetch(API_ENDPOINT, fetchOptions);
-  if (!res.ok) {
-    console.error(`Failed to fetch API: ${res.status} ${res.statusText}`);
-    if (res.status === 401) {
-      console.error("Invalid API key. Please check your VENICE_API_KEY.");
-    }
-    process.exit(1);
-  }
-
-  const json = await res.json();
-  const parsed = VeniceResponse.safeParse(json);
-  if (!parsed.success) {
-    console.error("Invalid API response:", parsed.error.errors);
-    process.exit(1);
-  }
-
-  const apiModels = parsed.data.data;
-
-  // Get existing files
-  const existingFiles = new Set<string>();
-  try {
-    const files = await readdir(modelsDir);
-    for (const file of files) {
-      if (file.endsWith(".toml")) {
-        existingFiles.add(file);
-      }
-    }
-  } catch {
-    // Directory might not exist yet
-  }
-
-  console.log(`Found ${apiModels.length} models in API, ${existingFiles.size} existing files\n`);
-
-  // Track API model IDs for orphan detection
-  const apiModelIds = new Set<string>();
-
-  let created = 0;
-  let updated = 0;
-  let unchanged = 0;
-
-  for (const apiModel of apiModels) {
-    const safeId = apiModel.id.replace(/\//g, "-");
-    const filename = `${safeId}.toml`;
-    const filePath = path.join(modelsDir, filename);
-
-    apiModelIds.add(filename);
-
-    const existing = await loadExistingModel(filePath);
-    const merged = mergeModel(apiModel, existing);
-    const tomlContent = formatToml(merged);
-
-    if (existing === null) {
-      // New file
-      created++;
-      if (dryRun) {
-        console.log(`[DRY RUN] Would create: ${filename}`);
-        console.log(`  name = "${merged.name}"`);
-        if (merged.family) {
-          console.log(`  family = "${merged.family}" (inferred)`);
-        }
-        console.log("");
-      } else {
-        await Bun.write(filePath, tomlContent);
-        console.log(`Created: ${filename}`);
-      }
-    } else {
-      // Check for changes
-      const changes = detectChanges(existing, merged);
-
-      if (changes.length > 0) {
-        updated++;
-        if (dryRun) {
-          console.log(`[DRY RUN] Would update: ${filename}`);
-        } else {
-          await Bun.write(filePath, tomlContent);
-          console.log(`Updated: ${filename}`);
-        }
-        for (const change of changes) {
-          console.log(`  ${change.field}: ${change.oldValue} → ${change.newValue}`);
-        }
-        console.log("");
-      } else {
-        unchanged++;
-      }
-    }
-  }
-
-  // Check for orphaned files
-  const orphaned: string[] = [];
-  for (const file of existingFiles) {
-    if (!apiModelIds.has(file)) {
-      orphaned.push(file);
-      console.log(`Warning: Orphaned file (not in API): ${file}`);
-    }
-  }
-
-  // Summary
-  console.log("");
-  if (dryRun) {
-    console.log(
-      `Summary: ${created} would be created, ${updated} would be updated, ${unchanged} unchanged, ${orphaned.length} orphaned`,
-    );
-  } else {
-    console.log(
-      `Summary: ${created} created, ${updated} updated, ${unchanged} unchanged, ${orphaned.length} orphaned`,
-    );
-  }
-}
-
-await main();

packages/core/script/generate-wandb.ts

@@ -3,7 +3,7 @@
 import path from "node:path";
 import { mkdir } from "node:fs/promises";
 import { z } from "zod";
-import { ModelFamilyValues } from "../src/family.js";
+import { inferKimiFamily, ModelFamilyValues } from "../src/family.js";
 
 const API_ENDPOINT = "https://trace.wandb.ai/inference/analysis/artificialanalysis/models";
 
@@ -176,6 +176,9 @@ function matchesFamily(target: string, family: string): boolean {
 }
 
 function inferFamily(modelId: string, modelName: string): string | undefined {
+  const kimiFamily = inferKimiFamily(modelId, modelName);
+  if (kimiFamily !== undefined) return kimiFamily;
+
   const sortedFamilies = [...ModelFamilyValues].sort((a, b) => b.length - a.length);
 
   for (const family of sortedFamilies) {

packages/core/src/family.ts

@@ -26,6 +26,7 @@ export const ModelFamilyValues = [
   "claude-haiku",
   "claude-sonnet",
   "claude-opus",
+  "claude-fable",
 
   // Gemini style
   "gemini",
@@ -65,8 +66,7 @@ export const ModelFamilyValues = [
 
   // Moonshot Kimi
   "kimi",
-  "kimi-k2.5",
-  "kimi-k2.6",
+  "kimi-k2",
   "kimi-free",
   "kimi-thinking",
 
@@ -102,6 +102,8 @@ export const ModelFamilyValues = [
   "command-r",
   "command-a",
   "command-light",
+  "north",
+  "north-free",
 
   // AI21 Jamba
   "jamba",
@@ -420,3 +422,10 @@ export const ModelFamilyValues = [
 
 export const ModelFamily = z.enum(ModelFamilyValues);
 export type ModelFamily = z.infer<typeof ModelFamily>;
+
+export function inferKimiFamily(...values: string[]): ModelFamily | undefined {
+  const target = values.join(" ").toLowerCase();
+  if (/kimi[^a-z0-9]*k2(?:[^a-z0-9]*\d+)?[^a-z0-9]*thinking/.test(target)) return "kimi-thinking";
+  if (/kimi[\s_-]*k2/.test(target)) return "kimi-k2";
+  return undefined;
+}

packages/core/src/schema.ts

@@ -25,7 +25,7 @@ const ReasoningEffortValue = z.preprocess(
   (value) => (value === "null" ? null : value),
   z.union([
     z.null(),
-    z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max"]),
+    z.enum(["none", "minimal", "low", "medium", "high", "xhigh", "max", "default"]),
   ]),
 );
 
@@ -47,7 +47,7 @@ const ReasoningOption = z
         type: z.literal("budget_tokens"),
         min: z
           .number()
-          .min(0, "Minimum reasoning budget cannot be negative")
+          .min(-1, "Minimum reasoning budget cannot be less than -1")
           .optional(),
         max: z
           .number()

packages/core/src/sync/index.ts

@@ -3,12 +3,14 @@ import { lstat, mkdir, readdir, rm } from "node:fs/promises";
 import { mergeDeep } from "remeda";
 import { z } from "zod";
 
-import { AuthoredModel, AuthoredModelShape } from "../schema.js";
+import { AuthoredModel, AuthoredModelShape, ModelMetadata } from "../schema.js";
+import { baseten } from "./providers/baseten.js";
 import { cloudflareWorkersAi } from "./providers/cloudflare-workers-ai.js";
 import { google } from "./providers/google.js";
 import { openrouter } from "./providers/openrouter.js";
 import { ovhcloud } from "./providers/ovhcloud.js";
 import { vercel } from "./providers/vercel.js";
+import { venice } from "./providers/venice.js";
 import { xai } from "./providers/xai.js";
 
 const ExistingModelType = AuthoredModelShape.partial()
@@ -39,14 +41,17 @@ export type ExistingModel = z.infer<typeof ExistingModelType>;
 export type SyncedFullModel = Omit<z.infer<typeof AuthoredModelShape>, "id">;
 export type SyncedBaseModel = Omit<z.infer<typeof SyncedBaseModel>, "id">;
 export type SyncedModel = SyncedFullModel | SyncedBaseModel;
+export type SyncedMetadata = Omit<z.infer<typeof ModelMetadata>, "id">;
 
 export interface SyncProvider<SourceModel> {
   id: string;
   name: string;
   modelsDir: string;
+  metadataNamespace?: string;
   skipCreates?: boolean;
   deleteMissing?: boolean;
   preserveSymlinks?: boolean;
+  preserveBaseModels?: boolean;
   sameModel?(current: ExistingModel, desired: SyncedModel): boolean;
   missingNotice?(paths: string[]): string[];
   sourceID?(model: SourceModel): string;
@@ -56,7 +61,7 @@ export interface SyncProvider<SourceModel> {
   translateModel(
     model: SourceModel,
     context: { existing(id: string): ExistingModel | undefined },
-  ): { id: string; model: SyncedModel } | undefined;
+  ): { id: string; model: SyncedModel; metadata?: { id: string; model: SyncedMetadata } } | undefined;
 }
 
 export interface SyncResult {
@@ -72,25 +77,29 @@ export interface SyncResult {
 }
 
 export const providers: {
+  baseten: SyncProvider<any>;
   "cloudflare-workers-ai": SyncProvider<any>;
   google: SyncProvider<any>;
   openrouter: SyncProvider<any>;
   ovhcloud: SyncProvider<any>;
   vercel: SyncProvider<any>;
+  venice: SyncProvider<any>;
   xai: SyncProvider<any>;
 } = {
+  baseten,
   "cloudflare-workers-ai": cloudflareWorkersAi,
   google,
   openrouter,
   ovhcloud,
   vercel,
+  venice,
   xai,
 };
 
 export const groups = {
   aggregators: ["openrouter", "vercel"],
   cloudflare: ["cloudflare-workers-ai"],
-  direct: ["google", "ovhcloud", "xai"],
+  direct: ["baseten", "google", "ovhcloud", "venice", "xai"],
 } as const;
 
 type ProviderID = keyof typeof providers;
@@ -110,9 +119,12 @@ export async function syncProvider<SourceModel>(
 ): Promise<SyncResult> {
   console.log(`\nSyncing ${provider.name}...`);
 
-  const { models: existing, brokenSymlinks } = await readExisting(provider.modelsDir);
+  const existingState = await readExisting(provider.modelsDir);
+  const { models: existing, brokenSymlinks } = existingState;
+  let { modelMetadata } = existingState;
   const sourceModels = provider.parseModels(await provider.fetchModels());
   const desired = new Map<string, { model: z.infer<typeof SyncedAuthoredModel>; content: string }>();
+  const desiredMetadata = new Map<string, { model: z.infer<typeof ModelMetadata>; content: string }>();
   const skippedRemote: string[] = [];
 
   for (const sourceModel of sourceModels) {
@@ -122,7 +134,7 @@ export async function syncProvider<SourceModel>(
       },
     });
     if (translated === undefined) {
-      if (provider.skipCreates) skippedRemote.push(provider.sourceID?.(sourceModel) ?? "unknown");
+      if (provider.sourceID !== undefined) skippedRemote.push(provider.sourceID(sourceModel));
       continue;
     }
 
@@ -136,9 +148,46 @@ export async function syncProvider<SourceModel>(
       throw new Error(`Duplicate synced model path: ${provider.id}/${relativePath}`);
     }
 
+    if (translated.metadata !== undefined) {
+      const parsedMetadata = ModelMetadata.safeParse({
+        id: translated.metadata.id,
+        ...stripUndefined(translated.metadata.model),
+      });
+      if (!parsedMetadata.success) {
+        parsedMetadata.error.cause = { provider: provider.id, metadata: translated.metadata.id };
+        throw parsedMetadata.error;
+      }
+      const metadataPath = `${translated.metadata.id}.toml`;
+      if (desiredMetadata.has(metadataPath)) throw new Error(`Duplicate synced metadata path: ${metadataPath}`);
+      desiredMetadata.set(metadataPath, {
+        model: parsedMetadata.data,
+        content: formatMetadataToml(parsedMetadata.data),
+      });
+    }
+
+    const translatedModel = provider.preserveBaseModels === false
+      ? translated.model
+      : preserveBaseModel(translated.model, existing.get(relativePath)?.authored);
+    const translatedBase = "base_model" in translatedModel ? translatedModel.base_model : undefined;
+    let resolvedReasoning: boolean | undefined;
+    if (translatedBase !== undefined) {
+      if (translated.metadata?.id === translatedBase) {
+        resolvedReasoning = translated.metadata.model.reasoning;
+      } else {
+        modelMetadata ??= await readModelMetadata(provider.modelsDir);
+        const canonicalReasoning = modelMetadata[translatedBase]?.reasoning;
+        resolvedReasoning = typeof canonicalReasoning === "boolean" ? canonicalReasoning : undefined;
+      }
+    } else {
+      resolvedReasoning = existing.get(relativePath)?.toml.reasoning;
+    }
     const parsed = SyncedAuthoredModel.safeParse(stripUndefined({
       id: translated.id,
-      ...preserveBaseModel(translated.model, existing.get(relativePath)?.authored),
+      ...preserveReasoningOptions(
+        translatedModel,
+        existing.get(relativePath)?.authored,
+        resolvedReasoning,
+      ),
     }));
     if (!parsed.success) {
       parsed.error.cause = { provider: provider.id, path: relativePath };
@@ -154,6 +203,48 @@ export async function syncProvider<SourceModel>(
   const files: SyncResult["files"] = [];
   let unchanged = 0;
 
+  const metadataDir = modelMetadataDir(provider.modelsDir);
+  for (const [relativePath, file] of desiredMetadata) {
+    const filePath = path.join(metadataDir, relativePath);
+    const currentFile = Bun.file(filePath);
+    const current = await currentFile.exists()
+      ? ModelMetadata.safeParse({
+          id: relativePath.slice(0, -5),
+          ...Bun.TOML.parse(await currentFile.text()) as Record<string, unknown>,
+        })
+      : undefined;
+    if (current?.success && stable(current.data) === stable(file.model)) continue;
+    files.push({ status: current === undefined ? "created" : "updated", path: filePath });
+    if (options.dryRun) {
+      console.log(`Would ${current === undefined ? "create" : "update"} metadata ${relativePath}`);
+    } else {
+      await mkdir(path.dirname(filePath), { recursive: true });
+      await Bun.write(filePath, file.content);
+    }
+  }
+
+  if (provider.metadataNamespace !== undefined) {
+    if (!/^[a-z0-9-]+$/.test(provider.metadataNamespace)) {
+      throw new Error(`Invalid metadata namespace: ${provider.metadataNamespace}`);
+    }
+    const namespaceDir = path.join(metadataDir, provider.metadataNamespace);
+    for (const { file } of await tomlFiles(namespaceDir)) {
+      const relativePath = path.join(provider.metadataNamespace, file);
+      if (desiredMetadata.has(relativePath) || provider.deleteMissing === false) continue;
+      if (options.newOnly) {
+        console.log(`Skipping metadata removal in new-only mode: ${relativePath}`);
+        continue;
+      }
+      const filePath = path.join(metadataDir, relativePath);
+      files.push({ status: "deleted", path: filePath });
+      if (options.dryRun) {
+        console.log(`Would remove metadata ${relativePath}`);
+      } else {
+        await rm(filePath, { force: true });
+      }
+    }
+  }
+
   for (const [relativePath, file] of desired) {
     const filePath = path.join(provider.modelsDir, relativePath);
     const current = existing.get(relativePath);
@@ -242,6 +333,22 @@ export function preserveBaseModel(model: SyncedModel, existing: ExistingModel |
   };
 }
 
+export function preserveReasoningOptions(
+  model: SyncedModel,
+  existing: ExistingModel | undefined,
+  resolvedReasoning: boolean | undefined = existing?.reasoning,
+): SyncedModel {
+  if ((model.reasoning ?? resolvedReasoning) === false) {
+    const { reasoning_options: _reasoningOptions, ...withoutReasoningOptions } = model;
+    return withoutReasoningOptions as SyncedModel;
+  }
+  if (model.reasoning_options !== undefined || existing?.reasoning_options === undefined) return model;
+  return {
+    ...model,
+    reasoning_options: existing.reasoning_options,
+  };
+}
+
 export async function syncTargets(target: string, options: SyncOptions = {}) {
   const ids = target in groups
     ? groups[target as keyof typeof groups]
@@ -307,7 +414,7 @@ async function readExisting(modelsDir: string) {
     existing.set(file, { authored, toml, symlink });
   }
 
-  return { models: existing, brokenSymlinks };
+  return { models: existing, brokenSymlinks, modelMetadata };
 }
 
 async function isSymlink(filePath: string) {
@@ -320,8 +427,7 @@ async function isSymlink(filePath: string) {
 }
 
 async function readModelMetadata(modelsDir: string) {
-  const root = path.dirname(path.dirname(path.dirname(modelsDir)));
-  const metadataDir = path.join(root, "models");
+  const metadataDir = modelMetadataDir(modelsDir);
   const result: Record<string, Record<string, unknown>> = {};
 
   for await (const modelPath of new Bun.Glob("**/*.toml").scan({
@@ -339,6 +445,10 @@ async function readModelMetadata(modelsDir: string) {
   return result;
 }
 
+function modelMetadataDir(modelsDir: string) {
+  return path.join(path.dirname(path.dirname(path.dirname(modelsDir))), "models");
+}
+
 function resolveBaseModel(
   authored: ExistingModel,
   modelMetadata: Record<string, Record<string, unknown>>,
@@ -585,15 +695,19 @@ function formatToml(model: z.infer<typeof SyncedAuthoredModel>) {
   if (model.open_weights !== undefined) lines.push(`open_weights = ${model.open_weights}`);
   if (model.status !== undefined) lines.push(`status = ${quote(model.status)}`);
 
-  for (const option of model.reasoning_options ?? []) {
-    lines.push("", "[[reasoning_options]]");
-    lines.push(`type = ${quote(option.type)}`);
-    if (option.type === "effort") {
-      lines.push(`values = [${option.values.map(formatReasoningValue).join(", ")}]`);
-    }
-    if (option.type === "budget_tokens") {
-      if (option.min !== undefined) lines.push(`min = ${formatInteger(option.min)}`);
-      if (option.max !== undefined) lines.push(`max = ${formatInteger(option.max)}`);
+  if (model.reasoning_options?.length === 0) {
+    lines.push("reasoning_options = []");
+  } else {
+    for (const option of model.reasoning_options ?? []) {
+      lines.push("", "[[reasoning_options]]");
+      lines.push(`type = ${quote(option.type)}`);
+      if (option.type === "effort") {
+        lines.push(`values = [${option.values.map(formatReasoningValue).join(", ")}]`);
+      }
+      if (option.type === "budget_tokens") {
+        if (option.min !== undefined) lines.push(`min = ${formatInteger(option.min)}`);
+        if (option.max !== undefined) lines.push(`max = ${formatInteger(option.max)}`);
+      }
     }
   }
 
@@ -658,6 +772,19 @@ function formatToml(model: z.infer<typeof SyncedAuthoredModel>) {
   return `${lines.join("\n")}\n`;
 }
 
+function formatMetadataToml(model: z.infer<typeof ModelMetadata>) {
+  const content = formatToml(model as unknown as z.infer<typeof SyncedAuthoredModel>).trimEnd();
+  const lines = [content];
+  for (const weight of model.weights ?? []) {
+    lines.push("", "[[weights]]");
+    if (weight.label !== undefined) lines.push(`label = ${quote(weight.label)}`);
+    lines.push(`url = ${quote(weight.url)}`);
+    if (weight.format !== undefined) lines.push(`format = ${quote(weight.format)}`);
+    if (weight.quantization !== undefined) lines.push(`quantization = ${quote(weight.quantization)}`);
+  }
+  return `${lines.join("\n")}\n`;
+}
+
 export async function main(args = process.argv.slice(2)) {
   if (args.includes("--list-providers")) {
     console.log(JSON.stringify(syncProviderMatrix()));

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

@@ -0,0 +1,188 @@
+import { z } from "zod";
+
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = "https://inference.baseten.co/v1/models";
+
+const Price = z.union([z.string(), z.number()]);
+
+export const BasetenModel = z.object({
+  id: z.string().min(1),
+  name: z.string().min(1),
+  context_length: z.number().int().positive(),
+  max_completion_tokens: z.number().int().positive(),
+  input_modalities: z.array(z.string()),
+  output_modalities: z.array(z.string()),
+  pricing: z.object({
+    prompt: Price,
+    completion: Price,
+  }).passthrough(),
+  supported_features: z.array(z.string()),
+  supported_sampling_parameters: z.array(z.string()),
+}).passthrough();
+
+export const BasetenResponse = z.object({
+  data: z.array(BasetenModel),
+}).passthrough();
+
+export type BasetenModel = z.infer<typeof BasetenModel>;
+
+export const baseten = {
+  id: "baseten",
+  name: "Baseten",
+  modelsDir: "providers/baseten/models",
+  deleteMissing: false,
+  sourceID(model) {
+    return model.id;
+  },
+  skippedNotice(ids) {
+    if (ids.length === 0) return [];
+    return [
+      `${ids.length} Baseten models were not created because their slugs could not be mapped exactly to provider-agnostic metadata.`,
+      `Skipped remote IDs: ${ids.map((id) => `\`${id}\``).join(", ")}`,
+    ];
+  },
+  missingNotice(paths) {
+    if (paths.length === 0) return [];
+    return [
+      `${paths.length} local Baseten models were absent from the catalog and were retained for manual lifecycle review.`,
+      `Retained local paths: ${paths.map((item) => `\`${item}\``).join(", ")}`,
+    ];
+  },
+  async fetchModels() {
+    const key = process.env.BASETEN_API_KEY;
+    if (key === undefined) throw new Error("Baseten sync requires BASETEN_API_KEY");
+    return fetchBasetenModels(key);
+  },
+  parseModels(raw) {
+    return BasetenResponse.parse(raw).data;
+  },
+  translateModel(model, context) {
+    const existing = context.existing(model.id);
+    const baseModel = existing === undefined
+      ? resolveBasetenBaseModel(model.id)
+      : existing.base_model;
+    if (existing === undefined && baseModel === undefined) return undefined;
+    if (
+      existing === undefined
+      && (price(model.pricing.prompt) === undefined || price(model.pricing.completion) === undefined)
+    ) return undefined;
+
+    return {
+      id: model.id,
+      model: buildBasetenModel(model, existing, baseModel),
+    };
+  },
+} satisfies SyncProvider<BasetenModel>;
+
+export async function fetchBasetenModels(
+  key: string,
+  fetcher: typeof fetch = fetch,
+) {
+  const response = await fetcher(API_ENDPOINT, {
+    headers: { Authorization: `Api-Key ${key}` },
+  });
+  if (!response.ok) {
+    throw new Error(`Baseten models request failed: ${response.status} ${response.statusText}`);
+  }
+  return BasetenResponse.parse(await response.json());
+}
+
+function price(value: string | number | undefined) {
+  if (value === undefined || value === "") return undefined;
+  const number = Number(value);
+  return Number.isFinite(number) && number >= 0
+    ? Math.round(number * 1_000_000_000_000) / 1_000_000
+    : undefined;
+}
+
+export function buildBasetenModel(
+  model: BasetenModel,
+  existing: ExistingModel | undefined,
+  baseModel = existing === undefined ? resolveBasetenBaseModel(model.id) : existing.base_model,
+): SyncedModel {
+  const features = new Set(model.supported_features);
+  const samplingParameters = new Set(model.supported_sampling_parameters);
+  const input = modalities(model.input_modalities, existing?.modalities?.input ?? ["text"]);
+  const output = modalities(model.output_modalities, existing?.modalities?.output ?? ["text"]);
+  const inputCost = price(model.pricing.prompt);
+  const outputCost = price(model.pricing.completion);
+  const cost = inputCost !== undefined && outputCost !== undefined
+    ? {
+        input: inputCost,
+        output: outputCost,
+        reasoning: existing?.cost?.reasoning,
+        cache_read: existing?.cost?.cache_read,
+        cache_write: existing?.cost?.cache_write,
+        tiers: existing?.cost?.tiers,
+      }
+    : existing?.cost;
+  const limit = {
+    context: model.context_length,
+    input: existing?.limit?.input,
+    output: model.max_completion_tokens,
+  };
+  const values: Partial<SyncedFullModel> = {
+    name: model.name ?? existing?.name,
+    family: existing?.family,
+    release_date: existing?.release_date,
+    last_updated: existing?.last_updated,
+    attachment: input.some((value) => value !== "text"),
+    reasoning: features.has("reasoning") || existing?.reasoning,
+    reasoning_options: existing?.reasoning_options,
+    temperature: samplingParameters.has("temperature"),
+    tool_call: features.has("tools") || existing?.tool_call,
+    structured_output: features.has("structured_outputs") || existing?.structured_output,
+    knowledge: existing?.knowledge,
+    open_weights: existing?.open_weights,
+    status: existing?.status,
+    interleaved: existing?.interleaved,
+    cost,
+    limit,
+    modalities: { input, output },
+  };
+
+  if (baseModel !== undefined) {
+    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 required = z.object({
+    name: z.string(),
+    release_date: z.string(),
+    last_updated: z.string(),
+    open_weights: z.boolean(),
+    cost: z.object({ input: z.number(), output: z.number() }),
+  }).safeParse(values);
+  if (!required.success) {
+    throw new Error(`Baseten model ${model.id} has incomplete local metadata required for sync`);
+  }
+  return values as SyncedFullModel;
+}
+
+export function resolveBasetenBaseModel(id: string) {
+  const [prefix, ...parts] = id.split("/");
+  if (prefix === undefined || parts.length === 0) return undefined;
+  const canonicalPrefix = {
+    "deepseek-ai": "deepseek",
+    MiniMaxAI: "minimax",
+    moonshotai: "moonshotai",
+    nvidia: "nvidia",
+    "zai-org": "zai",
+  }[prefix];
+  if (canonicalPrefix === undefined) return resolveCanonicalBaseModel(id);
+  return resolveCanonicalBaseModel(`${canonicalPrefix}/${parts.join("/").toLowerCase()}`);
+}
+
+type Modality = "text" | "audio" | "image" | "video" | "pdf";
+
+function modalities(values: string[], fallback: Modality[]): Modality[] {
+  const allowed = new Set<Modality>(["text", "audio", "image", "video", "pdf"]);
+  const result = values
+    .map((value) => value.toLowerCase())
+    .filter((value): value is Modality => allowed.has(value as Modality));
+  return [...new Set(result.length > 0 ? result : fallback)];
+}

packages/core/src/sync/providers/cloudflare-workers-ai.ts

@@ -96,7 +96,7 @@ export const cloudflareWorkersAi = {
   },
 } satisfies SyncProvider<CloudflareModel>;
 
-function buildWorkersAiModel(
+export function buildWorkersAiModel(
   model: z.infer<typeof OpenRouterModel>,
   existing: ExistingModel | undefined,
 ): SyncedModel {
@@ -108,11 +108,14 @@ function buildWorkersAiModel(
       max_completion_tokens: existing?.limit?.output ?? model.top_provider.max_completion_tokens,
     },
   };
-  const synced = buildOpenRouterModel(
-    source,
-    existing,
-    existing?.base_model ?? resolveCloudflareBaseModel(model),
-  );
+  const synced = {
+    ...buildOpenRouterModel(
+      source,
+      existing,
+      existing?.base_model ?? resolveCloudflareBaseModel(model),
+    ),
+    reasoning_options: existing?.reasoning_options,
+  };
   if ("base_model" in synced) return synced;
   return {
     ...synced,

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

@@ -1,6 +1,7 @@
 import { z } from "zod";
 
-import type { ExistingModel, SyncProvider, SyncedModel } from "../index.js";
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel } from "./openrouter.js";
 
 const API_ENDPOINT = "https://generativelanguage.googleapis.com/v1beta/models";
 
@@ -81,12 +82,12 @@ export const google = {
 
     return {
       id,
-      model: buildModel(model, existing),
+      model: buildGoogleModel(model, existing),
     };
   },
 } satisfies SyncProvider<GoogleModel>;
 
-function buildModel(model: GoogleModel, existing: ExistingModel): SyncedModel {
+export function buildGoogleModel(model: GoogleModel, existing: ExistingModel): SyncedModel {
   const name = existing.name;
   const releaseDate = existing.release_date;
   const lastUpdated = existing.last_updated;
@@ -111,9 +112,7 @@ function buildModel(model: GoogleModel, existing: ExistingModel): SyncedModel {
     throw new Error(`Google model ${model.name} has incomplete local TOML metadata required for sync`);
   }
 
-  return {
-    base_model: existing.base_model,
-    base_model_omit: existing.base_model_omit,
+  const synced: SyncedFullModel = {
     name: model.displayName ?? name,
     family: existing.family,
     release_date: releaseDate,
@@ -138,4 +137,8 @@ function buildModel(model: GoogleModel, existing: ExistingModel): SyncedModel {
     },
     modalities,
   };
+
+  return existing.base_model === undefined
+    ? synced
+    : factorBaseModel(existing.base_model, synced, synced.limit, existing.base_model_omit);
 }

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

@@ -2,7 +2,7 @@ import { z } from "zod";
 import { readFileSync, readdirSync } from "node:fs";
 import path from "node:path";
 
-import { ModelFamilyValues } from "../../family.js";
+import { inferKimiFamily, ModelFamilyValues } from "../../family.js";
 import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
 
 const API_ENDPOINT = "https://openrouter.ai/api/v1/models";
@@ -113,6 +113,9 @@ function modalities(values: string[], fallback: Modality[]): Modality[] {
 }
 
 function inferFamily(model: OpenRouterModel, name: string) {
+  const kimiFamily = inferKimiFamily(model.id, name);
+  if (kimiFamily !== undefined) return kimiFamily;
+
   const target = `${model.id} ${name}`.toLowerCase();
   return [...ModelFamilyValues]
     .sort((a, b) => b.length - a.length)
@@ -286,6 +289,14 @@ function baseModelOverrides(
 function inheritedOverride(value: unknown, inherited: unknown): unknown {
   if (value === undefined) return undefined;
   if (sameInheritedValue(value, inherited)) return undefined;
+  if (isPlainObject(value) && isPlainObject(inherited)) {
+    const overrides = Object.fromEntries(
+      Object.entries(value)
+        .map(([key, item]) => [key, inheritedOverride(item, inherited[key])])
+        .filter(([, item]) => item !== undefined),
+    );
+    return Object.keys(overrides).length > 0 ? overrides : undefined;
+  }
   return stripUndefined(value);
 }
 

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

@@ -121,6 +121,7 @@ export function buildOvhcloudModel(
     last_updated: lastUpdated,
     attachment,
     reasoning,
+    reasoning_options: reasoning ? existing?.reasoning_options : undefined,
     temperature: temperature || undefined,
     tool_call: toolCall,
     structured_output: structuredOutput || undefined,

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

@@ -0,0 +1,241 @@
+import { readdirSync } from "node:fs";
+import path from "node:path";
+import { z } from "zod";
+
+import { inferKimiFamily, ModelFamilyValues } from "../../family.js";
+import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
+import { factorBaseModel } from "./openrouter.js";
+
+const API_ENDPOINT = "https://api.venice.ai/api/v1/models?type=text";
+const MODELS_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..", "models");
+
+const Capabilities = z.object({
+  supportsAudioInput: z.boolean().optional(),
+  supportsE2EE: z.boolean().optional(),
+  supportsFunctionCalling: z.boolean().optional(),
+  supportsReasoning: z.boolean().optional(),
+  supportsReasoningEffort: z.boolean().optional(),
+  reasoningEffortOptions: z.array(z.string()).optional(),
+  supportsResponseSchema: z.boolean().optional(),
+  supportsVideoInput: z.boolean().optional(),
+  supportsVision: z.boolean().optional(),
+}).passthrough();
+
+const PricingTier = z.object({
+  usd: z.number().nonnegative(),
+}).passthrough();
+
+const ExtendedPricing = z.object({
+  context_token_threshold: z.number().int().nonnegative(),
+  input: PricingTier,
+  output: PricingTier,
+  cache_input: PricingTier.optional(),
+  cache_write: PricingTier.optional(),
+}).passthrough();
+
+const Pricing = z.object({
+  input: PricingTier,
+  output: PricingTier,
+  cache_input: PricingTier.optional(),
+  cache_write: PricingTier.optional(),
+  extended: ExtendedPricing.optional(),
+}).passthrough();
+
+const ModelSpec = z.object({
+  pricing: Pricing.optional(),
+  availableContextTokens: z.number().int().nonnegative(),
+  maxCompletionTokens: z.number().int().nonnegative().optional(),
+  capabilities: Capabilities,
+  name: z.string().min(1),
+  modelSource: z.string().optional(),
+}).passthrough();
+
+export const VeniceModel = z.object({
+  created: z.number(),
+  id: z.string().min(1),
+  model_spec: ModelSpec,
+}).passthrough();
+
+export const VeniceResponse = z.object({
+  data: z.array(VeniceModel),
+}).passthrough();
+
+export type VeniceModel = z.infer<typeof VeniceModel>;
+
+type ReasoningEffort = "default" | "max" | "low" | "high" | "none" | "medium" | "minimal" | "xhigh";
+
+interface MetadataEntry {
+  id: string;
+  filename: string;
+  normalizedFull: string;
+  normalizedFilename: string;
+}
+
+let metadataEntries: MetadataEntry[] | undefined;
+
+const BASE_MODEL_ALIASES: Record<string, string> = {
+  "claude-opus-4-6-fast": "anthropic/claude-opus-4-6",
+  "claude-opus-4-7-fast": "anthropic/claude-opus-4-7",
+  "claude-opus-4-8-fast": "anthropic/claude-opus-4-8",
+};
+
+export const venice = {
+  id: "venice",
+  name: "Venice",
+  modelsDir: "providers/venice/models",
+  preserveBaseModels: false,
+  async fetchModels() {
+    const headers = process.env.VENICE_API_KEY
+      ? { Authorization: `Bearer ${process.env.VENICE_API_KEY}` }
+      : undefined;
+    const response = await fetch(API_ENDPOINT, { headers });
+    if (!response.ok) {
+      throw new Error(`Venice models request failed: ${response.status} ${response.statusText}`);
+    }
+    return response.json();
+  },
+  parseModels(raw) {
+    return VeniceResponse.parse(raw).data;
+  },
+  translateModel(model, context) {
+    if (model.model_spec.capabilities.supportsE2EE === true) return undefined;
+    const id = model.id.replaceAll("/", "-");
+    const existing = context.existing(id);
+    const existingBase = existing?.base_model?.startsWith("venice/") === false ? existing.base_model : undefined;
+    const resolvedBase = existingBase ?? resolveVeniceBaseModel(model.id, model.model_spec.name);
+    return {
+      id,
+      model: buildVeniceModel(model, existing, resolvedBase ?? null),
+    };
+  },
+} satisfies SyncProvider<VeniceModel>;
+
+export function buildVeniceModel(
+  model: VeniceModel,
+  existing: ExistingModel | undefined,
+  baseModel: string | null | undefined = existing?.base_model ?? resolveVeniceBaseModel(model.id, model.model_spec.name),
+  today = new Date().toISOString().slice(0, 10),
+): SyncedModel {
+  const spec = model.model_spec;
+  const capabilities = spec.capabilities;
+  const input = [
+    "text" as const,
+    ...(capabilities.supportsVision ? ["image" as const] : []),
+    ...(capabilities.supportsAudioInput ? ["audio" as const] : []),
+    ...(capabilities.supportsVideoInput ? ["video" as const] : []),
+    ...(existing?.modalities?.input.includes("pdf") ? ["pdf" as const] : []),
+  ];
+  const limit = {
+    context: spec.availableContextTokens,
+    input: existing?.limit?.input,
+    output: spec.maxCompletionTokens ?? Math.floor(spec.availableContextTokens / 4),
+  };
+  const reasoningEfforts = capabilities.reasoningEffortOptions?.filter(isReasoningEffort);
+  const reasoningOptions = reasoningEfforts?.length
+    ? [{ type: "effort" as const, values: reasoningEfforts }]
+    : [];
+  const cost = spec.pricing === undefined
+    ? existing?.cost
+    : {
+        input: spec.pricing.input.usd,
+        output: spec.pricing.output.usd,
+        reasoning: existing?.cost?.reasoning,
+        cache_read: spec.pricing.cache_input?.usd,
+        cache_write: spec.pricing.cache_write?.usd,
+        input_audio: existing?.cost?.input_audio,
+        output_audio: existing?.cost?.output_audio,
+        tiers: spec.pricing.extended === undefined
+          ? existing?.cost?.tiers
+          : [{
+              tier: { type: "context" as const, size: spec.pricing.extended.context_token_threshold },
+              input: spec.pricing.extended.input.usd,
+              output: spec.pricing.extended.output.usd,
+              cache_read: spec.pricing.extended.cache_input?.usd,
+              cache_write: spec.pricing.extended.cache_write?.usd,
+            }],
+      };
+  const authoritative = {
+    name: spec.name,
+    attachment: input.some((value) => value !== "text"),
+    reasoning: capabilities.supportsReasoning === true,
+    reasoning_options: reasoningOptions,
+    tool_call: capabilities.supportsFunctionCalling === true,
+    structured_output: capabilities.supportsResponseSchema === true ? true : undefined,
+    temperature: undefined,
+    cost,
+    limit,
+    modalities: { input: [...new Set(input)], output: ["text" as const] },
+  };
+  const releaseDate = new Date(model.created * 1000).toISOString().slice(0, 10);
+  const values: SyncedFullModel = {
+    ...authoritative,
+    family: baseModel == null ? inferFamily(model.id, spec.name) ?? existing?.family : existing?.family,
+    release_date: releaseDate,
+    last_updated: existing?.last_updated ?? today,
+    knowledge: existing?.knowledge,
+    open_weights: spec.modelSource?.toLowerCase().includes("huggingface")
+      ?? existing?.open_weights
+      ?? false,
+    status: existing?.status,
+    interleaved: existing?.interleaved,
+  };
+
+  return baseModel == null
+    ? values
+    : factorBaseModel(baseModel, values, limit, existing?.base_model_omit);
+}
+
+export function resolveVeniceBaseModel(id: string, name: string) {
+  const alias = BASE_MODEL_ALIASES[id];
+  if (alias !== undefined) return alias;
+  const entries = getMetadataEntries();
+  const normalizedID = normalize(id);
+  const normalizedName = normalize(name);
+  const ranked = [
+    entries.filter((entry) => entry.normalizedFull === normalizedID),
+    entries.filter((entry) => entry.normalizedFilename === normalizedID),
+    entries.filter((entry) => entry.normalizedFilename === normalizedName),
+  ];
+  return ranked.find((matches) => matches.length === 1)?.[0]?.id;
+}
+
+function getMetadataEntries() {
+  if (metadataEntries !== undefined) return metadataEntries;
+  metadataEntries = [];
+  for (const provider of readdirSync(MODELS_DIR, { withFileTypes: true })) {
+    if (!provider.isDirectory()) continue;
+    for (const file of readdirSync(path.join(MODELS_DIR, provider.name), { withFileTypes: true })) {
+      if (!file.isFile() || !file.name.endsWith(".toml")) continue;
+      const filename = file.name.slice(0, -5);
+      metadataEntries.push({
+        id: `${provider.name}/${filename}`,
+        filename,
+        normalizedFull: normalize(`${provider.name}/${filename}`),
+        normalizedFilename: normalize(filename),
+      });
+    }
+  }
+  return metadataEntries;
+}
+
+function normalize(value: string) {
+  return value.toLowerCase().replaceAll(/[^a-z0-9]/g, "");
+}
+
+function isReasoningEffort(value: string): value is ReasoningEffort {
+  return ["default", "max", "low", "high", "none", "medium", "minimal", "xhigh"].includes(value);
+}
+
+function inferFamily(id: string, name: string) {
+  const kimiFamily = inferKimiFamily(id, name);
+  if (kimiFamily !== undefined) return kimiFamily;
+
+  const target = `${id} ${name}`.toLowerCase();
+  return [...ModelFamilyValues]
+    .sort((a, b) => b.length - a.length)
+    .find((family) => {
+      const value = family.toLowerCase().replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+      if (family === "o") return new RegExp(`(^|[^a-z0-9])${value}(?=\\d|$|[^a-z0-9])`).test(target);
+      return new RegExp(`(^|[^a-z0-9])${value}(?=$|[^a-z0-9])`).test(target);
+    });
+}

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

@@ -1,6 +1,6 @@
 import { z } from "zod";
 
-import { ModelFamilyValues } from "../../family.js";
+import { inferKimiFamily, ModelFamilyValues } from "../../family.js";
 import type { ExistingModel, SyncProvider, SyncedFullModel, SyncedModel } from "../index.js";
 import { factorBaseModel, resolveCanonicalBaseModel } from "./openrouter.js";
 
@@ -47,11 +47,7 @@ export const vercel = {
   id: "vercel",
   name: "Vercel AI Gateway",
   modelsDir: "providers/vercel/models",
-  deleteMissing: false,
   preserveSymlinks: true,
-  missingNotice(paths) {
-    return paths.map((model) => `Vercel model is no longer returned by the API: ${model}`);
-  },
   async fetchModels() {
     const response = await fetch(API_ENDPOINT);
     if (!response.ok) {
@@ -63,9 +59,6 @@ export const vercel = {
     return VercelResponse.parse(raw).data;
   },
   translateModel(model, context) {
-    if (model.type === "image" || model.type === "video" || model.type === "reranking") {
-      return undefined;
-    }
     return {
       id: model.id,
       model: buildVercelModel(model, context.existing(model.id)),
@@ -101,7 +94,9 @@ export function buildVercelModel(model: VercelModel, existing: ExistingModel | u
     reasoning: existing?.reasoning ?? tags.has("reasoning"),
     reasoning_options: existing?.reasoning_options,
     temperature: true,
-    tool_call: existing?.tool_call ?? tags.has("tool-use"),
+    tool_call: model.type === "language"
+      ? existing?.tool_call ?? tags.has("tool-use")
+      : tags.has("tool-use"),
     structured_output: existing?.structured_output,
     knowledge: existing?.knowledge,
     open_weights: existing?.open_weights ?? false,
@@ -114,7 +109,13 @@ export function buildVercelModel(model: VercelModel, existing: ExistingModel | u
     modalities: {
       input: ["text", tags.has("vision") ? "image" : undefined, tags.has("file-input") ? "pdf" : undefined]
         .filter((value): value is "text" | "image" | "pdf" => value !== undefined),
-      output: tags.has("image-generation") ? ["text", "image"] : ["text"],
+      output: model.type === "image"
+        ? ["image"]
+        : model.type === "video"
+        ? ["video"]
+        : tags.has("image-generation")
+        ? ["text", "image"]
+        : ["text"],
     },
   };
 
@@ -152,6 +153,9 @@ function buildCost(pricing: VercelModel["pricing"], existing?: ExistingModel["co
 }
 
 function inferFamily(modelID: string, name: string) {
+  const kimiFamily = inferKimiFamily(modelID, name);
+  if (kimiFamily !== undefined) return kimiFamily;
+
   const targets = [modelID, name].map((value) => value.toLowerCase());
   const families = [...ModelFamilyValues].sort((a, b) => b.length - a.length);
   return families.find((family) => targets.some((target) => target.includes(family.toLowerCase())))
@@ -175,6 +179,7 @@ function sameVercelModel(current: ExistingModel, desired: SyncedModel) {
     [current.family, desiredModel.family],
     [current.attachment, desiredModel.attachment],
     [current.reasoning, desiredModel.reasoning],
+    [current.reasoning_options, desiredModel.reasoning_options],
     [current.tool_call, desiredModel.tool_call],
     [current.structured_output, desiredModel.structured_output],
     [current.open_weights, desiredModel.open_weights],

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

@@ -29,7 +29,7 @@ const XAIAPIKey = z.object({
   acls: z.array(z.string()),
 }).passthrough();
 
-type XAIModel = z.infer<typeof XAIModel>;
+export type XAIModel = z.infer<typeof XAIModel>;
 
 export const xai = {
   id: "xai",
@@ -87,7 +87,7 @@ export const xai = {
 
     return {
       id: model.id,
-      model: buildModel(model, existing),
+      model: buildXAIModel(model, existing),
     };
   },
 } satisfies SyncProvider<XAIModel>;
@@ -159,7 +159,7 @@ function cost(model: XAIModel, existing: ExistingModel) {
   };
 }
 
-function buildModel(model: XAIModel, existing: ExistingModel): SyncedModel {
+export function buildXAIModel(model: XAIModel, existing: ExistingModel): SyncedModel {
   const name = existing.name;
   const attachment = existing.attachment;
   const reasoning = existing.reasoning;
@@ -195,6 +195,7 @@ function buildModel(model: XAIModel, existing: ExistingModel): SyncedModel {
     last_updated: model.canonical_id === undefined ? created : lastUpdated!,
     attachment: input.some((value) => value !== "text"),
     reasoning,
+    reasoning_options: existing.reasoning_options,
     temperature: existing.temperature,
     tool_call: toolCall,
     structured_output: existing.structured_output,

packages/core/test/baseten-sync.test.ts

@@ -0,0 +1,150 @@
+import { afterEach, expect, test } from "bun:test";
+import path from "node:path";
+import { mkdir, mkdtemp } from "node:fs/promises";
+import os from "node:os";
+
+import { syncProvider } from "../src/sync/index.js";
+import {
+  BasetenResponse,
+  baseten,
+  buildBasetenModel,
+  fetchBasetenModels,
+  type BasetenModel,
+} from "../src/sync/providers/baseten.js";
+
+const catalogModel: BasetenModel = {
+  id: "zai-org/GLM-5.1",
+  name: "GLM 5.1",
+  context_length: 128_000,
+  max_completion_tokens: 32_000,
+  input_modalities: ["text"],
+  output_modalities: ["text"],
+  pricing: {
+    prompt: "0.00000012",
+    completion: "0.0000005",
+  },
+  supported_features: ["reasoning", "reasoning_effort", "tools", "structured_outputs"],
+  supported_sampling_parameters: ["temperature", "top_p"],
+};
+
+const newCatalogModel: BasetenModel = {
+  ...catalogModel,
+};
+
+afterEach(() => {
+  baseten.modelsDir = "providers/baseten/models";
+  baseten.fetchModels = async () => {
+    const key = process.env.BASETEN_API_KEY;
+    if (key === undefined) throw new Error("Baseten sync requires BASETEN_API_KEY");
+    return fetchBasetenModels(key);
+  };
+});
+
+test("Baseten maps authoritative fields and preserves curated metadata", () => {
+  const synced = buildBasetenModel(catalogModel, {
+    name: "Old name",
+    release_date: "2025-08-05",
+    last_updated: "2025-09-01",
+    attachment: false,
+    reasoning: true,
+    reasoning_options: [{ type: "effort", values: ["low", "high"] }],
+    tool_call: true,
+    open_weights: true,
+    status: "deprecated",
+    interleaved: { field: "reasoning_content" },
+    base_model: "zhipuai/glm-5.1",
+    base_model_omit: ["limit.input"],
+    cost: { input: 0.1, output: 0.4, cache_write: 0.2 },
+    limit: { context: 64_000, output: 16_000 },
+    modalities: { input: ["text"], output: ["text"] },
+  });
+
+  expect(synced).toMatchObject({
+    base_model: "zhipuai/glm-5.1",
+    base_model_omit: ["limit.input"],
+    reasoning_options: [{ type: "effort", values: ["low", "high"] }],
+    status: "deprecated",
+    interleaved: { field: "reasoning_content" },
+    cost: { input: 0.12, output: 0.5, cache_write: 0.2 },
+    limit: { context: 128_000, output: 32_000 },
+  });
+});
+
+test("Baseten preserves curated reasoning when an opt-in capability is omitted", () => {
+  const synced = buildBasetenModel({
+    ...catalogModel,
+    supported_features: ["tools", "structured_outputs"],
+  }, {
+    name: "GLM 5.1",
+    release_date: "2026-05-20",
+    last_updated: "2026-05-20",
+    attachment: false,
+    reasoning: true,
+    reasoning_options: [{ type: "toggle" }],
+    tool_call: true,
+    open_weights: true,
+    cost: { input: 1, output: 4 },
+    limit: { context: 100_000, output: 50_000 },
+    modalities: { input: ["text"], output: ["text"] },
+  });
+
+  expect(synced).toMatchObject({
+    reasoning: true,
+    reasoning_options: [{ type: "toggle" }],
+  });
+});
+
+test("Baseten sync adds exact base models, retains missing entries, and is idempotent", async () => {
+  const root = await mkdtemp(path.join(os.tmpdir(), "models-dev-baseten-"));
+  const modelsDir = path.join(root, "providers", "baseten", "models");
+  const metadataDir = path.join(root, "models", "zhipuai");
+  await mkdir(path.join(modelsDir, "stale"), { recursive: true });
+  await mkdir(metadataDir, { recursive: true });
+  await Bun.write(
+    path.join(metadataDir, "glm-5.1.toml"),
+    Bun.file(path.join(import.meta.dirname, "../../../models/zhipuai/glm-5.1.toml")),
+  );
+  await Bun.write(path.join(modelsDir, "stale", "model.toml"), [
+    'name = "Retained"',
+    'release_date = "2025-01-01"',
+    'last_updated = "2025-01-01"',
+    "attachment = false",
+    "reasoning = false",
+    "tool_call = false",
+    "open_weights = false",
+    "[cost]",
+    "input = 1",
+    "output = 1",
+    "[limit]",
+    "context = 1000",
+    "output = 100",
+    "[modalities]",
+    'input = ["text"]',
+    'output = ["text"]',
+    "",
+  ].join("\n"));
+  baseten.modelsDir = modelsDir;
+  baseten.fetchModels = async () => ({ data: [newCatalogModel] });
+
+  const first = await syncProvider(baseten);
+  const second = await syncProvider(baseten);
+
+  expect(first.created).toBe(1);
+  expect(first.deleted).toBe(0);
+  expect(first.notices.join(" ")).toContain("stale/model.toml");
+  expect(second).toMatchObject({ created: 0, updated: 0, deleted: 0 });
+});
+
+test("Baseten rejects malformed catalog responses", () => {
+  expect(() => BasetenResponse.parse({ data: "broken" })).toThrow();
+});
+
+test("Baseten rejects non-success API responses", async () => {
+  const fetcher = async () => new Response("unauthorized", {
+    status: 401,
+    statusText: "Unauthorized",
+  });
+
+  expect(fetchBasetenModels("fixture-key", fetcher as typeof fetch))
+    .rejects.toThrow("Baseten models request failed: 401 Unauthorized");
+});

packages/core/test/cloudflare-workers-ai-sync.test.ts

@@ -0,0 +1,35 @@
+import { expect, test } from "bun:test";
+
+import { buildWorkersAiModel } from "../src/sync/providers/cloudflare-workers-ai.js";
+import type { OpenRouterModel } from "../src/sync/providers/openrouter.js";
+
+test("Cloudflare Workers AI sync preserves reasoning options", () => {
+  const model: OpenRouterModel = {
+    id: "@cf/nvidia/nemotron-3-120b-a12b",
+    name: "Nemotron 3 Super 120B",
+    created: 1_773_187_200,
+    hugging_face_id: "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16",
+    knowledge_cutoff: null,
+    context_length: 256_000,
+    architecture: {
+      input_modalities: ["text"],
+      output_modalities: ["text"],
+    },
+    pricing: {
+      prompt: "0.0000005",
+      completion: "0.0000015",
+    },
+    top_provider: {
+      context_length: 256_000,
+      max_completion_tokens: 256_000,
+    },
+    supported_parameters: ["reasoning", "tools", "temperature"],
+  };
+
+  const synced = buildWorkersAiModel(model, {
+    base_model: "nvidia/nemotron-3-super-120b-a12b",
+    reasoning_options: [{ type: "toggle" }],
+  });
+
+  expect(synced.reasoning_options).toEqual([{ type: "toggle" }]);
+});

packages/core/test/family.test.ts

@@ -0,0 +1,15 @@
+import { expect, test } from "bun:test";
+
+import { inferKimiFamily } from "../src/family.js";
+
+test("Kimi family inference ignores K2 versions", () => {
+  expect(inferKimiFamily("moonshotai/kimi-k2.5")).toBe("kimi-k2");
+  expect(inferKimiFamily("moonshotai/kimi-k2.7-code")).toBe("kimi-k2");
+  expect(inferKimiFamily("Kimi K2.6")).toBe("kimi-k2");
+});
+
+test("Kimi family inference preserves thinking variants", () => {
+  expect(inferKimiFamily("moonshotai/kimi-k2-thinking")).toBe("kimi-thinking");
+  expect(inferKimiFamily("Kimi K2.5 Thinking")).toBe("kimi-thinking");
+  expect(inferKimiFamily("moonshotai/kimi-k2.6:thinking")).toBe("kimi-thinking");
+});

packages/core/test/google-sync.test.ts

@@ -0,0 +1,35 @@
+import { expect, test } from "bun:test";
+
+import { buildGoogleModel } from "../src/sync/providers/google.js";
+
+test("Google sync keeps base models compact", () => {
+  const synced = buildGoogleModel({
+    name: "models/gemini-3-pro-image-preview",
+    displayName: "Nano Banana Pro",
+    inputTokenLimit: 131_072,
+    outputTokenLimit: 32_768,
+    temperature: 1,
+    thinking: true,
+  }, {
+    base_model: "google/gemini-3-pro-image-preview",
+    name: "Nano Banana Pro",
+    family: "gemini-pro",
+    release_date: "2025-11-20",
+    last_updated: "2025-11-20",
+    attachment: true,
+    reasoning: true,
+    temperature: true,
+    tool_call: false,
+    knowledge: "2025-01",
+    open_weights: false,
+    cost: { input: 2, output: 120 },
+    limit: { context: 65_536, output: 32_768 },
+    modalities: { input: ["text", "image"], output: ["text", "image"] },
+  });
+
+  expect(synced).toEqual({
+    base_model: "google/gemini-3-pro-image-preview",
+    cost: { input: 2, output: 120 },
+    limit: { context: 131_072 },
+  });
+});

packages/core/test/sync-runner.test.ts

@@ -3,6 +3,7 @@ import path from "node:path";
 import { mkdtemp, mkdir, readlink, symlink } from "node:fs/promises";
 import os from "node:os";
 
+import { AuthoredModelShape } from "../src/schema.js";
 import { syncProvider, type SyncProvider, type SyncedFullModel } from "../src/sync/index.js";
 
 const model: SyncedFullModel = {
@@ -18,6 +19,28 @@ const model: SyncedFullModel = {
   modalities: { input: ["text"], output: ["text"] },
 };
 
+test("reasoning budgets allow only the -1 negative sentinel", () => {
+  const authored = { id: "model", ...model };
+
+  expect(AuthoredModelShape.safeParse({
+    ...authored,
+    reasoning_options: [{ type: "budget_tokens", min: -1, max: 32_768 }],
+  }).success).toBe(true);
+  expect(AuthoredModelShape.safeParse({
+    ...authored,
+    reasoning_options: [{ type: "budget_tokens", min: -2, max: 32_768 }],
+  }).success).toBe(false);
+});
+
+test("reasoning efforts accept the provider default value", () => {
+  expect(AuthoredModelShape.safeParse({
+    id: "model",
+    ...model,
+    reasoning: true,
+    reasoning_options: [{ type: "effort", values: ["none", "default"] }],
+  }).success).toBe(true);
+});
+
 async function fixture() {
   const root = await mkdtemp(path.join(os.tmpdir(), "models-dev-sync-"));
   const modelsDir = path.join(root, "providers", "test", "models");
@@ -97,3 +120,108 @@ test("non-deleting sync reports missing broken symlinks", async () => {
   expect(result.notices).toEqual(["missing: model.toml"]);
   expect(await readlink(filePath)).toBe("missing.toml");
 });
+
+test("sync preserves authored reasoning options omitted by a translator", async () => {
+  const { modelsDir } = await fixture();
+  const filePath = path.join(modelsDir, "model.toml");
+  await Bun.write(filePath, `name = "Old name"
+release_date = "2026-01-01"
+last_updated = "2026-01-01"
+attachment = false
+reasoning = true
+tool_call = false
+open_weights = false
+
+[[reasoning_options]]
+type = "effort"
+values = ["low", "high"]
+
+[[reasoning_options]]
+type = "budget_tokens"
+min = -1
+max = 32768
+
+[cost]
+input = 1
+output = 2
+
+[limit]
+context = 1000
+output = 100
+
+[modalities]
+input = ["text"]
+output = ["text"]
+`);
+  const sync = provider(modelsDir, ["model"]);
+  sync.translateModel = (id) => ({
+    id,
+    model: { ...model, reasoning: true },
+  });
+
+  const first = await syncProvider(sync);
+  const content = await Bun.file(filePath).text();
+  const second = await syncProvider(sync);
+
+  expect(first.updated).toBe(1);
+  expect(content).toContain("[[reasoning_options]]");
+  expect(content).toContain('values = ["low", "high"]');
+  expect(content).toContain("min = -1");
+  expect(content).toContain("max = 32_768");
+  expect(second.updated).toBe(0);
+  expect(second.unchanged).toBe(1);
+});
+
+test("sync writes metadata returned by a provider translator", async () => {
+  const root = await mkdtemp(path.join(os.tmpdir(), "models-dev-sync-metadata-"));
+  const modelsDir = path.join(root, "providers", "test", "models");
+  await mkdir(modelsDir, { recursive: true });
+  const sync = provider(modelsDir, ["model"]);
+  sync.translateModel = () => ({
+    id: "model",
+    model: {
+      base_model: "test/model",
+      reasoning_options: [],
+      cost: { input: 1, output: 2 },
+    },
+    metadata: {
+      id: "test/model",
+      model: {
+        name: "Model",
+        release_date: "2026-06-10",
+        last_updated: "2026-06-10",
+        attachment: false,
+        reasoning: false,
+        tool_call: true,
+        open_weights: false,
+        limit: { context: 1_000, output: 100 },
+        modalities: { input: ["text"], output: ["text"] },
+      },
+    },
+  });
+
+  const first = await syncProvider(sync);
+  const second = await syncProvider(sync);
+
+  expect(first).toMatchObject({ created: 2, updated: 0 });
+  expect(second).toMatchObject({ created: 0, updated: 0 });
+  expect(await Bun.file(path.join(root, "models", "test", "model.toml")).text()).toContain('name = "Model"');
+});
+
+test("sync removes missing metadata only from its owned namespace", async () => {
+  const { root, modelsDir } = await fixture();
+  const ownedDir = path.join(root, "models", "test");
+  const otherDir = path.join(root, "models", "other");
+  await mkdir(ownedDir, { recursive: true });
+  await mkdir(otherDir, { recursive: true });
+  await Bun.write(path.join(ownedDir, "stale.toml"), 'name = "Stale"\n');
+  await Bun.write(path.join(otherDir, "retained.toml"), 'name = "Retained"\n');
+  const sync = provider(modelsDir, []);
+  sync.metadataNamespace = "test";
+
+  const result = await syncProvider(sync);
+
+  expect(result.deleted).toBe(1);
+  expect(await Bun.file(path.join(ownedDir, "stale.toml")).exists()).toBe(false);
+  expect(await Bun.file(path.join(otherDir, "retained.toml")).exists()).toBe(true);
+});

packages/core/test/venice-sync.test.ts

@@ -0,0 +1,167 @@
+import { expect, test } from "bun:test";
+import { readdirSync } from "node:fs";
+import path from "node:path";
+
+import {
+  buildVeniceModel,
+  resolveVeniceBaseModel,
+  venice,
+  VeniceResponse,
+  type VeniceModel,
+} from "../src/sync/providers/venice.js";
+
+const catalogModel: VeniceModel = {
+  id: "openai-gpt-54",
+  created: 1_772_668_800,
+  model_spec: {
+    name: "GPT-5.4",
+    availableContextTokens: 400_000,
+    maxCompletionTokens: 128_000,
+    modelSource: "OpenAI",
+    capabilities: {
+      supportsVision: true,
+      supportsReasoning: true,
+      supportsReasoningEffort: true,
+      reasoningEffortOptions: ["none", "low", "medium", "high"],
+      supportsFunctionCalling: true,
+      supportsResponseSchema: true,
+    },
+    pricing: {
+      input: { usd: 3.13 },
+      output: { usd: 18.75 },
+      cache_input: { usd: 0.313 },
+      extended: {
+        context_token_threshold: 200_000,
+        input: { usd: 6.26 },
+        output: { usd: 28.125 },
+      },
+    },
+  },
+};
+
+test("Venice resolves flattened IDs to canonical metadata", () => {
+  expect(resolveVeniceBaseModel("openai-gpt-54", "GPT-5.4")).toBe("openai/gpt-5.4");
+  expect(resolveVeniceBaseModel("claude-opus-4-8-fast", "Claude Opus 4.8 Fast"))
+    .toBe("anthropic/claude-opus-4-8");
+});
+
+test("Venice emits empty reasoning options when efforts are unavailable", () => {
+  const synced = buildVeniceModel({
+    ...catalogModel,
+    id: "reasoning-without-efforts",
+    model_spec: {
+      ...catalogModel.model_spec,
+      name: "Reasoning Without Efforts",
+      capabilities: {
+        ...catalogModel.model_spec.capabilities,
+        reasoningEffortOptions: [],
+      },
+    },
+  }, undefined, undefined, "2026-06-10");
+
+  expect(synced).toMatchObject({ reasoning: true, reasoning_options: [] });
+});
+
+test("Venice does not infer temperature support", () => {
+  const synced = buildVeniceModel(catalogModel, undefined, null, "2026-06-10");
+
+  expect(synced.temperature).toBeUndefined();
+});
+
+test("Venice skips E2EE models", () => {
+  const translated = venice.translateModel({
+    ...catalogModel,
+    id: "e2ee-test-model",
+    model_spec: {
+      ...catalogModel.model_spec,
+      capabilities: { ...catalogModel.model_spec.capabilities, supportsE2EE: true },
+    },
+  }, { existing: () => undefined });
+
+  expect(translated).toBeUndefined();
+});
+
+test("Venice uses boundary-aware family matching", () => {
+  const synced = buildVeniceModel({
+    ...catalogModel,
+    id: "google-gemma-4-31b-it",
+    model_spec: { ...catalogModel.model_spec, name: "Google Gemma 4 31B Instruct" },
+  }, undefined, null, "2026-06-10");
+
+  expect(synced).toMatchObject({ family: "gemma" });
+});
+
+test("Venice maps API fields without bumping inherited model timestamps", () => {
+  const synced = buildVeniceModel(catalogModel, {
+    base_model: "openai/gpt-5.4",
+    name: "GPT-5.4",
+    family: "gpt",
+    release_date: "2026-03-05",
+    last_updated: "2026-03-09",
+    attachment: true,
+    reasoning: true,
+    tool_call: true,
+    structured_output: true,
+    temperature: true,
+    open_weights: false,
+    interleaved: { field: "reasoning_content" },
+    cost: { input: 3, output: 18, input_audio: 4 },
+    limit: { context: 400_000, output: 128_000 },
+    modalities: { input: ["text", "image", "pdf"], output: ["text"] },
+  }, "openai/gpt-5.4", "2026-06-10");
+
+  expect(synced).toMatchObject({
+    base_model: "openai/gpt-5.4",
+    base_model_omit: ["limit.input"],
+    last_updated: "2026-03-09",
+    reasoning_options: [{ type: "effort", values: ["none", "low", "medium", "high"] }],
+    interleaved: { field: "reasoning_content" },
+    cost: {
+      input: 3.13,
+      output: 18.75,
+      cache_read: 0.313,
+      input_audio: 4,
+      tiers: [{ tier: { type: "context", size: 200_000 }, input: 6.26, output: 28.125 }],
+    },
+  });
+  expect(synced).not.toHaveProperty("family");
+  expect(synced).not.toHaveProperty("release_date");
+  expect(synced).not.toHaveProperty("open_weights");
+  expect(synced).not.toHaveProperty("modalities");
+  expect(synced).not.toHaveProperty("temperature");
+});
+
+test("Venice preserves last_updated when authoritative data is unchanged", () => {
+  const providerModel = {
+    ...catalogModel,
+    id: "venice-only-test-model",
+    model_spec: { ...catalogModel.model_spec, name: "Venice Only Test Model" },
+  };
+  const full = buildVeniceModel(providerModel, undefined, undefined, "2026-06-10");
+  if ("base_model" in full) throw new Error("Expected a full provider model fixture");
+  const synced = buildVeniceModel(providerModel, full, undefined, "2026-06-11");
+
+  expect(synced).toMatchObject({ last_updated: "2026-06-10" });
+});
+
+test("Venice rejects malformed responses", () => {
+  expect(() => VeniceResponse.parse({ data: [{ id: "broken" }] })).toThrow();
+});
+
+test("Venice models use only canonical metadata and declare reasoning options", async () => {
+  const root = path.join(import.meta.dirname, "..", "..", "..");
+  const modelsDir = path.join(root, "providers", "venice", "models");
+
+  for (const file of readdirSync(modelsDir).filter((item) => item.endsWith(".toml"))) {
+    const model = Bun.TOML.parse(await Bun.file(path.join(modelsDir, file)).text()) as {
+      base_model?: string;
+      reasoning_options?: unknown[];
+    };
+    expect(model.reasoning_options, file).toBeDefined();
+    if (model.base_model !== undefined) {
+      expect(model.base_model.startsWith("venice/"), file).toBe(false);
+      expect(await Bun.file(path.join(root, "models", `${model.base_model}.toml`)).exists(), file).toBe(true);
+    }
+    expect(file.startsWith("e2ee-"), file).toBe(false);
+  }
+});

packages/core/test/vercel-sync.test.ts

@@ -1,6 +1,6 @@
 import { expect, test } from "bun:test";
 
-import { buildVercelModel, type VercelModel } from "../src/sync/providers/vercel.js";
+import { buildVercelModel, type VercelModel, vercel } from "../src/sync/providers/vercel.js";
 
 const model: VercelModel = {
   id: "openai/gpt-test",
@@ -64,6 +64,42 @@ test("Vercel models preserve curated metadata and missing limits", () => {
   expect(synced.limit).toEqual({ context: 64_000, input: 48_000, output: 16_000 });
 });
 
+test("Vercel non-language models use API tool capabilities", () => {
+  const synced = buildVercelModel({
+    ...model,
+    type: "image",
+    tags: [],
+  }, {
+    tool_call: true,
+  });
+
+  expect(synced.tool_call).toBe(false);
+});
+
+test("Vercel sync includes non-language model types", () => {
+  for (const [type, output] of [
+    ["image", ["image"]],
+    ["video", ["video"]],
+    ["reranking", ["text"]],
+  ] as const) {
+    const source = {
+      ...model,
+      id: `test/${type}`,
+      type,
+      tags: [],
+      context_window: 0,
+      max_tokens: 0,
+      pricing: undefined,
+    };
+
+    expect(vercel.translateModel(source, { existing: () => undefined })).toBeDefined();
+    expect(buildVercelModel(source, undefined)).toMatchObject({
+      tool_call: false,
+      modalities: { input: ["text"], output },
+    });
+  }
+});
+
 test("Vercel models use canonical metadata when available", () => {
   const synced = buildVercelModel({
     ...model,

packages/core/test/xai-sync.test.ts

@@ -0,0 +1,30 @@
+import { expect, test } from "bun:test";
+
+import { buildXAIModel, type XAIModel } from "../src/sync/providers/xai.js";
+
+const model: XAIModel = {
+  id: "grok-test",
+  created: 1_700_000_000,
+  input_modalities: ["text"],
+  output_modalities: ["text"],
+  prompt_text_token_price: 10_000,
+  completion_text_token_price: 20_000,
+};
+
+test("xAI sync preserves reasoning options", () => {
+  const synced = buildXAIModel(model, {
+    name: "Grok Test",
+    release_date: "2024-01-01",
+    last_updated: "2024-01-01",
+    attachment: false,
+    reasoning: true,
+    reasoning_options: [{ type: "effort", values: ["none", "high"] }],
+    tool_call: true,
+    open_weights: false,
+    limit: { context: 2_000_000, output: 30_000 },
+    modalities: { input: ["text"], output: ["text"] },
+  });
+
+  expect(synced.reasoning_options).toEqual([{ type: "effort", values: ["none", "high"] }]);
+  expect(synced.limit?.context).toBe(2_000_000);
+});

providers/abacus/models/Qwen/QwQ-32B.toml

@@ -4,6 +4,7 @@ release_date = "2024-11-28"
 last_updated = "2024-11-28"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/Qwen/Qwen3-235B-A22B-Instruct-2507.toml

@@ -4,6 +4,7 @@ release_date = "2025-07-01"
 last_updated = "2025-07-01"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/Qwen/Qwen3-32B.toml

@@ -4,6 +4,7 @@ release_date = "2025-04-29"
 last_updated = "2025-04-29"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/Qwen/qwen3-coder-480b-a35b-instruct.toml

@@ -4,6 +4,7 @@ release_date = "2025-07-22"
 last_updated = "2025-07-22"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/claude-3-7-sonnet-20250219.toml

@@ -4,6 +4,7 @@ release_date = "2025-02-19"
 last_updated = "2025-02-19"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2024-10-31"

providers/abacus/models/claude-haiku-4-5-20251001.toml

@@ -4,6 +4,7 @@ release_date = "2025-10-15"
 last_updated = "2025-10-15"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-02-28"

providers/abacus/models/claude-opus-4-1-20250805.toml

@@ -4,6 +4,7 @@ release_date = "2025-08-05"
 last_updated = "2025-08-05"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = false

providers/abacus/models/claude-opus-4-20250514.toml

@@ -4,6 +4,7 @@ release_date = "2025-05-14"
 last_updated = "2025-05-14"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = false

providers/abacus/models/claude-opus-4-5-20251101.toml

@@ -4,6 +4,7 @@ release_date = "2025-11-01"
 last_updated = "2025-11-01"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-03-31"

providers/abacus/models/claude-opus-4-6.toml

@@ -4,6 +4,7 @@ release_date = "2026-02-05"
 last_updated = "2026-02-05"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-05-31"

providers/abacus/models/claude-sonnet-4-20250514.toml

@@ -4,6 +4,7 @@ release_date = "2025-05-14"
 last_updated = "2025-05-14"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = false

providers/abacus/models/claude-sonnet-4-5-20250929.toml

@@ -4,6 +4,7 @@ release_date = "2025-09-29"
 last_updated = "2025-09-29"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-07-31"

providers/abacus/models/claude-sonnet-4-6.toml

@@ -4,6 +4,7 @@ release_date = "2026-02-17"
 last_updated = "2026-02-17"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-08-31"

providers/abacus/models/deepseek-ai/DeepSeek-R1.toml

@@ -4,6 +4,7 @@ release_date = "2025-01-20"
 last_updated = "2025-01-20"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/deepseek-ai/DeepSeek-V3.1-Terminus.toml

@@ -4,6 +4,7 @@ release_date = "2025-06-01"
 last_updated = "2025-06-01"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/deepseek-ai/DeepSeek-V3.2.toml

@@ -4,6 +4,7 @@ release_date = "2025-06-15"
 last_updated = "2025-06-15"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/deepseek/deepseek-v3.1.toml

@@ -4,6 +4,7 @@ release_date = "2025-01-20"
 last_updated = "2025-01-20"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/gemini-2.5-flash.toml

@@ -4,6 +4,7 @@ release_date = "2025-03-20"
 last_updated = "2025-06-05"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-01"

providers/abacus/models/gemini-2.5-pro.toml

@@ -4,6 +4,7 @@ release_date = "2025-03-25"
 last_updated = "2025-03-25"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-01"

providers/abacus/models/gemini-3-flash-preview.toml

@@ -4,6 +4,7 @@ release_date = "2025-12-17"
 last_updated = "2025-12-17"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 knowledge = "2025-01"

providers/abacus/models/gemini-3.1-flash-lite-preview.toml

@@ -4,6 +4,7 @@ release_date = "2026-03-01"
 last_updated = "2026-03-01"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 structured_output = true

providers/abacus/models/gemini-3.1-pro-preview.toml

@@ -4,6 +4,7 @@ release_date = "2026-02-19"
 last_updated = "2026-02-19"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 structured_output = true

providers/abacus/models/gpt-5-codex.toml

@@ -4,6 +4,7 @@ release_date = "2025-09-15"
 last_updated = "2025-09-15"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/gpt-5-mini.toml

@@ -4,6 +4,7 @@ release_date = "2025-08-07"
 last_updated = "2025-08-07"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-05-30"
 tool_call = true

providers/abacus/models/gpt-5-nano.toml

@@ -4,6 +4,7 @@ release_date = "2025-08-07"
 last_updated = "2025-08-07"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-05-30"
 tool_call = true

providers/abacus/models/gpt-5.1-chat-latest.toml

@@ -4,6 +4,7 @@ release_date = "2025-11-13"
 last_updated = "2025-11-13"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/gpt-5.1-codex-max.toml

@@ -4,6 +4,7 @@ release_date = "2025-11-13"
 last_updated = "2025-11-13"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/gpt-5.1-codex.toml

@@ -4,6 +4,7 @@ release_date = "2025-11-13"
 last_updated = "2025-11-13"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/gpt-5.1.toml

@@ -4,6 +4,7 @@ release_date = "2025-11-13"
 last_updated = "2025-11-13"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/gpt-5.2-chat-latest.toml

@@ -4,6 +4,7 @@ release_date = "2026-01-01"
 last_updated = "2026-01-01"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/gpt-5.2-codex.toml

@@ -4,6 +4,7 @@ release_date = "2025-12-11"
 last_updated = "2025-12-11"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2025-08-31"
 tool_call = true

providers/abacus/models/gpt-5.2.toml

@@ -4,6 +4,7 @@ release_date = "2025-12-11"
 last_updated = "2025-12-11"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2025-08-31"
 tool_call = true

providers/abacus/models/gpt-5.3-chat-latest.toml

@@ -4,6 +4,7 @@ release_date = "2026-03-01"
 last_updated = "2026-03-01"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = false

providers/abacus/models/gpt-5.3-codex-xhigh.toml

@@ -4,6 +4,7 @@ release_date = "2026-02-05"
 last_updated = "2026-02-05"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2025-08-31"
 tool_call = true

providers/abacus/models/gpt-5.3-codex.toml

@@ -4,6 +4,7 @@ release_date = "2026-02-05"
 last_updated = "2026-02-05"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2025-08-31"
 tool_call = true

providers/abacus/models/gpt-5.4.toml

@@ -4,6 +4,7 @@ release_date = "2026-03-05"
 last_updated = "2026-03-05"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2025-08-31"
 tool_call = true

providers/abacus/models/gpt-5.toml

@@ -4,6 +4,7 @@ release_date = "2025-08-07"
 last_updated = "2025-08-07"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-09-30"
 tool_call = true

providers/abacus/models/grok-4-0709.toml

@@ -4,6 +4,7 @@ release_date = "2025-07-09"
 last_updated = "2025-07-09"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = false

providers/abacus/models/kimi-k2-turbo-preview.toml

@@ -1,5 +1,5 @@
 name = "Kimi K2 Turbo Preview"
-family = "kimi"
+family = "kimi-k2"
 release_date = "2025-07-08"
 last_updated = "2025-07-08"
 attachment = false

providers/abacus/models/kimi-k2.5.toml

@@ -1,9 +1,10 @@
 name = "Kimi K2.5"
-family = "kimi"
+family = "kimi-k2"
 release_date = "2026-01"
 last_updated = "2026-01"
 attachment = false
 reasoning = true
+reasoning_options = []
 structured_output = true
 temperature = true
 tool_call = true

providers/abacus/models/o3-mini.toml

@@ -4,6 +4,7 @@ release_date = "2024-12-20"
 last_updated = "2025-01-29"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-05"
 tool_call = true

providers/abacus/models/o3-pro.toml

@@ -4,6 +4,7 @@ release_date = "2025-06-10"
 last_updated = "2025-06-10"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-05"
 tool_call = true

providers/abacus/models/o3.toml

@@ -4,6 +4,7 @@ release_date = "2025-04-16"
 last_updated = "2025-04-16"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-05"
 tool_call = true

providers/abacus/models/o4-mini.toml

@@ -4,6 +4,7 @@ release_date = "2025-04-16"
 last_updated = "2025-04-16"
 attachment = true
 reasoning = true
+reasoning_options = []
 temperature = false
 knowledge = "2024-05"
 tool_call = true

providers/abacus/models/openai/gpt-oss-120b.toml

@@ -4,6 +4,7 @@ release_date = "2025-08-05"
 last_updated = "2025-08-05"
 attachment = true
 reasoning = true
+reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/qwen3-max.toml

@@ -4,6 +4,7 @@ release_date = "2025-05-28"
 last_updated = "2025-05-28"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = false

providers/abacus/models/zai-org/glm-4.5.toml

@@ -4,6 +4,7 @@ release_date = "2025-07-28"
 last_updated = "2025-07-28"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/abacus/models/zai-org/glm-5.toml

@@ -4,6 +4,7 @@ release_date = "2026-02-11"
 last_updated = "2026-02-11"
 attachment = false
 reasoning = true
+reasoning_options = []
 temperature = true
 tool_call = true
 open_weights = true

providers/aihubmix/models/coding-xiaomi-mimo-v2.5-pro.toml

@@ -1,4 +1,5 @@
 base_model = "xiaomi/mimo-v2.5-pro"
+reasoning_options = [{ type = "toggle" }]
 name = "Coding Xiaomi MiMo-V2.5-Pro"
 family = "mimo-v2.5-pro"
 last_updated = "2026-05-13"

providers/aihubmix/models/coding-xiaomi-mimo-v2.5.toml

@@ -1,4 +1,5 @@
 base_model = "xiaomi/mimo-v2.5"
+reasoning_options = [{ type = "toggle" }]
 name = "Coding Xiaomi MiMo-V2.5"
 family = "mimo-v2.5"
 last_updated = "2026-05-13"

providers/aihubmix/models/kimi-k2.5.toml

@@ -1,5 +1,5 @@
 name = "Kimi K2.5"
-family = "kimi-k2.5"
+family = "kimi-k2"
 release_date = "2026-01"
 last_updated = "2026-01"
 attachment = true

providers/aihubmix/models/kimi-k2.6.toml

@@ -1,5 +1,5 @@
 name = "Kimi K2.6"
-family = "kimi-k2.6"
+family = "kimi-k2"
 release_date = "2026-04-21"
 last_updated = "2026-04-21"
 attachment = true

providers/aihubmix/models/xiaomi-mimo-v2.5-free.toml

@@ -1,4 +1,5 @@
 base_model = "xiaomi/mimo-v2.5"
+reasoning_options = [{ type = "toggle" }]
 name = "Xiaomi MiMo-V2.5 (free)"
 family = "mimo-v2.5"
 last_updated = "2026-05-13"