free-coding-models 0.5.60 → 0.5.62

package/src/core/models-drift.js

@@ -0,0 +1,296 @@
+/**
+ * @file models-drift.js
+ * @description Drift detection between `sources.js` (our curated catalog) and the live
+ *              models.dev community catalog. Compares context window, max output tokens,
+ *              and capability flags (reasoning / vision / thinking). Reports mismatches
+ *              but never auto-rewrites sources.js — a human approves the edit.
+ *
+ * @details
+ *   📖 Why this exists:
+ *   📖 - sources.js is curated but `ctx` drifts constantly (128k → 256k → 1M).
+ *   📖 - We don't want to silently overwrite curated values; we want to surface
+ *   📖   the drift so a human can review it.
+ *   📖 - Used by:
+ *   📖     1. `scripts/check-drift.mjs` (CLI: `free-coding-models --check-drift`)
+ *   📖     2. `.github/workflows/check-drift.yml` (weekly CI — opens an issue)
+ *   📖     3. The TUI footer chip + /health endpoint (read-only)
+ *
+ *   📖 What "drift" means here:
+ *   📖   - ctx: sources.js says "128k", models.dev says 256000 → UPDATE
+ *   📖   - maxTokens: sources.js has no value, models.dev says 65536 → ADD
+ *   📖   - reasoning/vision flags: sources.js is null, models.dev has the flag → ADD
+ *   📖   - exact match: ✓ (no action)
+ *
+ *   📖 A "field mismatch" entry has:
+ *   📖   { modelId, field, sourcesJsValue, modelsDevValue, action: 'update'|'add'|'remove' }
+ *
+ * @functions
+ *   → detectDrift(models, catalog, opts?)              — Returns a list of field mismatches
+ *   → summarizeDrift(mismatches)                       — { total, byField, byModel, modelsAffected }
+ *   → formatDriftReport(mismatches, opts?)             — Pretty-print a human-readable report
+ *   → parseCtxToNum(ctx)                               — "128k" / "1m" → number
+ *   → DRIFT_FIELDS                                     — The list of fields we check
+ *
+ * @exports detectDrift, summarizeDrift, formatDriftReport, parseCtxToNum, DRIFT_FIELDS
+ *
+ * @see src/core/models-dev-fetcher.js — provides the live catalog
+ * @see src/core/models-dev-index.js  — provides the lookup helpers
+ * @see scripts/check-drift.mjs       — CLI consumer
+ */
+
+import { buildModelIndex, lookupModelDevMeta, normalizeModelDevEntry } from './models-dev-index.js'
+
+/** 📖 Fields we check for drift. Each has a sources.js column + a models.dev mapping. */
+export const DRIFT_FIELDS = ['ctx', 'maxTokens', 'reasoning', 'vision', 'thinking']
+
+/**
+ * 📖 Convert a sources.js ctx string ("128k", "1m", "262144") to a number.
+ * 📖 Mirrors parseCtxToK from utils.js but returns raw tokens (not thousands).
+ * 📖 Returns null if the string is empty, "-", or unparseable.
+ */
+export function parseCtxToNum(ctx) {
+  if (ctx == null) return null
+  const s = String(ctx).trim()
+  if (!s || s === '-' || s === '—') return null
+  // Pure number
+  if (/^\d+$/.test(s)) return parseInt(s, 10)
+  // Suffix
+  const m = s.match(/^(\d+(?:\.\d+)?)\s*([km])$/i)
+  if (m) {
+    const n = parseFloat(m[1])
+    const unit = m[2].toLowerCase()
+    if (unit === 'k') return Math.round(n * 1000)
+    if (unit === 'm') return Math.round(n * 1_000_000)
+  }
+  return null
+}
+
+// ─── Comparison helpers ──────────────────────────────────────────────────────
+
+function contextMatch(sourcesJsCtx, devContextWindow) {
+  const a = parseCtxToNum(sourcesJsCtx)
+  const b = (typeof devContextWindow === 'number' && Number.isFinite(devContextWindow)) ? devContextWindow : null
+  if (a == null && b == null) return { status: 'both-null' }
+  if (a == null && b != null) return { status: 'add', sourcesJs: null, dev: b }
+  if (a != null && b == null) return { status: 'dev-missing', sourcesJs: a, dev: null }
+  // 📖 Allow 5% tolerance for vendor-specific rounding
+  if (Math.abs(a - b) <= Math.max(a, b) * 0.05) return { status: 'match', sourcesJs: a, dev: b }
+  return { status: 'drift', sourcesJs: a, dev: b }
+}
+
+function flagMatch(sourcesJsFlag, devFlag) {
+  // 📖 sources.js stores capability flags as booleans (true/false) or null/undefined
+  const a = sourcesJsFlag === true
+  const b = devFlag === true
+  if (!a && !b) return { status: 'both-off' }
+  if (!a && b) return { status: 'add', sourcesJs: false, dev: true }
+  if (a && !b) return { status: 'drift', sourcesJs: true, dev: false }
+  return { status: 'match' }
+}
+
+function numMatch(sourcesJsNum, devNum) {
+  const a = (typeof sourcesJsNum === 'number' && Number.isFinite(sourcesJsNum)) ? sourcesJsNum : null
+  const b = (typeof devNum === 'number' && Number.isFinite(devNum)) ? devNum : null
+  if (a == null && b == null) return { status: 'both-null' }
+  if (a == null && b != null) return { status: 'add', sourcesJs: null, dev: b }
+  if (a != null && b == null) return { status: 'dev-missing', sourcesJs: a, dev: null }
+  if (a === b) return { status: 'match', sourcesJs: a, dev: b }
+  return { status: 'drift', sourcesJs: a, dev: b }
+}
+
+// ─── Main detection ──────────────────────────────────────────────────────────
+
+/**
+ * 📖 Run drift detection over the sources.js model list against the live catalog.
+ * 📖 Returns a list of per-field mismatches. Use `summarizeDrift` + `formatDriftReport`
+ * 📖 to render the result.
+ *
+ * 📖 Each model in `models` should be a sources.js tuple:
+ * 📖   [modelId, label, tier, sweScore, ctx, providerKey, ...]
+ * 📖 or an object with the same fields.
+ *
+ * @param {Array} models — sources.js MODELS array
+ * @param {object} catalog — parsed models.dev catalog (or null = no drift)
+ * @param {object} [opts]
+ * @param {object} [opts.index] — Pre-built index (skips the build)
+ * @param {number} [opts.threshold=0] — Min mismatches to report; 0 = all
+ * @returns {Array<{
+ *   modelId: string, label: string, field: string,
+ *   sourcesJsValue: any, modelsDevValue: any,
+ *   action: 'update'|'add'|'drift', matchKind: 'exact'|'alias'|'substring'|'none'
+ * }>}
+ */
+export function detectDrift(models, catalog, opts = {}) {
+  if (!Array.isArray(models) || models.length === 0) return []
+  if (!catalog || typeof catalog !== 'object') return []
+
+  const index = opts.index ?? buildModelIndex(catalog)
+  const threshold = opts.threshold ?? 0
+  const out = []
+
+  for (const entry of models) {
+    // 📖 Accept both tuple form and object form
+    let modelId, label, ctx, reasoning, vision, thinking, maxTokens
+    if (Array.isArray(entry)) {
+      [modelId, label, , , ctx] = entry
+      // 📖 The 6th+ elements of the sources.js tuple are providerKey + metadata;
+      // 📖 for drift we only need the first 5.
+    } else if (entry && typeof entry === 'object') {
+      ({ modelId, label, ctx, reasoning, vision, thinking, maxTokens } = entry)
+    } else {
+      continue
+    }
+    if (!modelId) continue
+
+    const match = lookupModelDevMeta(modelId, label, index)
+    if (!match) continue
+    const norm = match.entry
+    const matchKind = match.matchKind
+
+    // 📖 Skip substring matches for ctx/metadata to avoid false positives.
+    // 📖 Only exact + alias matches contribute to drift.
+    if (matchKind === 'substring') continue
+
+    // ctx
+    const ctxCmp = contextMatch(ctx, norm.contextWindow)
+    if (ctxCmp.status === 'drift' || ctxCmp.status === 'add') {
+      out.push({
+        modelId, label, field: 'ctx',
+        sourcesJsValue: ctxCmp.sourcesJs,
+        modelsDevValue: ctxCmp.dev,
+        action: ctxCmp.status === 'add' ? 'add' : 'update',
+        matchKind,
+      })
+    }
+    // maxTokens
+    const maxCmp = numMatch(maxTokens, norm.maxOutputTokens)
+    if (maxCmp.status === 'drift' || maxCmp.status === 'add') {
+      out.push({
+        modelId, label, field: 'maxTokens',
+        sourcesJsValue: maxCmp.sourcesJs,
+        modelsDevValue: maxCmp.dev,
+        action: maxCmp.status === 'add' ? 'add' : 'update',
+        matchKind,
+      })
+    }
+    // reasoning
+    const rCmp = flagMatch(reasoning, norm.reasoning)
+    if (rCmp.status === 'drift' || rCmp.status === 'add') {
+      out.push({
+        modelId, label, field: 'reasoning',
+        sourcesJsValue: rCmp.sourcesJs,
+        modelsDevValue: rCmp.dev,
+        action: rCmp.status === 'add' ? 'add' : 'update',
+        matchKind,
+      })
+    }
+    // vision
+    const vCmp = flagMatch(vision, norm.vision)
+    if (vCmp.status === 'drift' || vCmp.status === 'add') {
+      out.push({
+        modelId, label, field: 'vision',
+        sourcesJsValue: vCmp.sourcesJs,
+        modelsDevValue: vCmp.dev,
+        action: vCmp.status === 'add' ? 'add' : 'update',
+        matchKind,
+      })
+    }
+    // thinking
+    const tCmp = flagMatch(thinking, norm.thinking)
+    if (tCmp.status === 'drift' || tCmp.status === 'add') {
+      out.push({
+        modelId, label, field: 'thinking',
+        sourcesJsValue: tCmp.sourcesJs,
+        modelsDevValue: tCmp.dev,
+        action: tCmp.status === 'add' ? 'add' : 'update',
+        matchKind,
+      })
+    }
+  }
+
+  if (threshold > 0 && out.length < threshold) {
+    return []
+  }
+  return out
+}
+
+// ─── Summary + report ────────────────────────────────────────────────────────
+
+/**
+ * 📖 Aggregate stats over a drift result list. Used by the TUI footer chip and
+ * 📖 the /health endpoint.
+ *
+ * @param {Array} mismatches — Output of detectDrift
+ * @returns {{
+ *   total: number,
+ *   byField: Record<string, number>,
+ *   byAction: Record<string, number>,
+ *   modelsAffected: string[]
+ * }}
+ */
+export function summarizeDrift(mismatches) {
+  const byField = Object.fromEntries(DRIFT_FIELDS.map(f => [f, 0]))
+  const byAction = { update: 0, add: 0, drift: 0 }
+  const modelsSet = new Set()
+  for (const m of mismatches || []) {
+    byField[m.field] = (byField[m.field] ?? 0) + 1
+    byAction[m.action] = (byAction[m.action] ?? 0) + 1
+    modelsSet.add(m.modelId)
+  }
+  return {
+    total: (mismatches || []).length,
+    byField,
+    byAction,
+    modelsAffected: Array.from(modelsSet).sort(),
+  }
+}
+
+/**
+ * 📖 Format a drift report for human reading. Used by the CLI script and the
+ * 📖 GitHub Actions workflow (which posts the report to an issue).
+ *
+ * @param {Array} mismatches — Output of detectDrift
+ * @param {object} [opts]
+ * @param {boolean} [opts.useColor=true] — ANSI-color the output
+ * @returns {string} Multi-line report
+ */
+export function formatDriftReport(mismatches, opts = {}) {
+  const useColor = opts.useColor !== false && process.stdout?.isTTY === true
+  const RED = useColor ? '\x1b[31m' : ''
+  const YEL = useColor ? '\x1b[33m' : ''
+  const GRN = useColor ? '\x1b[32m' : ''
+  const DIM = useColor ? '\x1b[2m' : ''
+  const RST = useColor ? '\x1b[0m' : ''
+
+  const summary = summarizeDrift(mismatches)
+  if (summary.total === 0) {
+    return `${GRN}✓ No catalog drift detected${RST} ${DIM}(all models match models.dev)${RST}`
+  }
+  const lines = []
+  lines.push(`${YEL}⚠️  Catalog drift detected (${summary.total} mismatches across ${summary.modelsAffected.length} models)${RST}`)
+  lines.push('')
+
+  // 📖 Group by model for readability
+  const byModel = new Map()
+  for (const m of mismatches) {
+    const list = byModel.get(m.modelId) ?? []
+    list.push(m)
+    byModel.set(m.modelId, list)
+  }
+  for (const [modelId, list] of byModel) {
+    const label = list[0].label || modelId
+    lines.push(`  ${label} ${DIM}(${modelId})${RST}`)
+    for (const m of list) {
+      const arrow = m.action === 'add' ? '← ADD' : '← UPDATE'
+      const color = m.action === 'add' ? GRN : RED
+      const sj = m.sourcesJsValue === null || m.sourcesJsValue === undefined ? '—' : String(m.sourcesJsValue)
+      const dv = m.modelsDevValue === null || m.modelsDevValue === undefined ? '—' : String(m.modelsDevValue)
+      lines.push(`    ${DIM}${m.field.padEnd(11)}${RST} sources.js=${sj}  models.dev=${dv}  ${color}${arrow}${RST}`)
+    }
+    lines.push('')
+  }
+  return lines.join('\n')
+}
+
+// Re-export normalizeModelDevEntry for convenience (drift callers often need it)
+export { normalizeModelDevEntry }

package/src/core/utils.js

@@ -606,6 +606,18 @@ export function parseArgs(argv) {
   const probeTtlRaw = probeTtlValueIdx !== -1 ? args[probeTtlValueIdx] : null
   const probeTtlMs = probeTtlRaw !== null ? parseInt(probeTtlRaw, 10) : null
 
+  // 📖 Drift detection flags (t5):
+  // 📖 --check-drift (boolean) — print a drift report vs models.dev, exit non-zero on mismatch.
+  // 📖 --drift-threshold <N> (value) — only fail when N+ mismatches are found.
+  const checkDriftMode = flags.includes('--check-drift')
+  const driftThresholdIdx = args.findIndex(a => a.toLowerCase() === '--drift-threshold')
+  const driftThresholdValueIdx = (driftThresholdIdx !== -1 && args[driftThresholdIdx + 1] && !args[driftThresholdIdx + 1].startsWith('--'))
+    ? driftThresholdIdx + 1
+    : -1
+  if (driftThresholdValueIdx !== -1) skipIndices.add(driftThresholdValueIdx)
+  const driftThresholdRaw = driftThresholdValueIdx !== -1 ? args[driftThresholdValueIdx] : null
+  const driftThreshold = driftThresholdRaw !== null ? parseInt(driftThresholdRaw, 10) : null
+
   return {
     apiKey,
     bestMode,
@@ -647,6 +659,8 @@ export function parseArgs(argv) {
     daemonMode,
     daemonBackgroundMode,
     daemonStopMode,
+    checkDriftMode,
+    driftThreshold,
     daemonStatusMode,
     // 📖 Profile system removed - API keys now persist permanently across all sessions
     recommendMode,

package/src/data/benchmarks.json

@@ -0,0 +1,302 @@
+{
+  "_meta": {
+    "schemaVersion": 1,
+    "lastUpdated": "2026-07-25",
+    "source": "Artificial Analysis + curated manual overlay",
+    "notes": "Curated seed dataset for ~30 well-known models. Refreshed by scripts/update-benchmarks.mjs before each release. Fields with null are not yet measured."
+  },
+  "deepseek-ai/deepseek-v3.2": {
+    "codingIndex": 65.4, "mathIndex": 70.2, "agenticIndex": 58.0, "reasoningIndex": 68.5,
+    "mmluPro": 75.3, "gpqa": 52.0, "hle": 8.4, "contextWindow": 160000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek V3.2"
+  },
+  "deepseek-ai/deepseek-v4-pro": {
+    "codingIndex": 81.2, "mathIndex": 84.5, "agenticIndex": 74.8, "reasoningIndex": 80.1,
+    "mmluPro": 86.4, "gpqa": 65.7, "hle": 14.2, "contextWindow": 1000000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek V4 Pro"
+  },
+  "deepseek-ai/deepseek-v4-flash": {
+    "codingIndex": 78.9, "mathIndex": 80.1, "agenticIndex": 71.5, "reasoningIndex": 77.0,
+    "mmluPro": 83.2, "gpqa": 60.4, "hle": 12.8, "contextWindow": 1000000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek V4 Flash"
+  },
+  "deepseek-ai/deepseek-r1": {
+    "codingIndex": 52.0, "mathIndex": 90.4, "agenticIndex": 48.5, "reasoningIndex": 78.6,
+    "mmluPro": 84.0, "gpqa": 71.5, "hle": 9.9, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek R1"
+  },
+  "deepseek/deepseek-r1": {
+    "codingIndex": 52.0, "mathIndex": 90.4, "agenticIndex": 48.5, "reasoningIndex": 78.6,
+    "mmluPro": 84.0, "gpqa": 71.5, "hle": 9.9, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek R1"
+  },
+  "DeepSeek-V3.1": {
+    "codingIndex": 68.0, "mathIndex": 73.0, "agenticIndex": 60.5, "reasoningIndex": 70.0,
+    "mmluPro": 76.5, "gpqa": 53.2, "hle": 9.1, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek V3.1"
+  },
+  "DeepSeek-V3.2": {
+    "codingIndex": 70.5, "mathIndex": 75.0, "agenticIndex": 62.0, "reasoningIndex": 72.0,
+    "mmluPro": 78.0, "gpqa": 54.5, "hle": 9.5, "contextWindow": 32000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek V3.2"
+  },
+  "deepseek/deepseek-chat": {
+    "codingIndex": 65.0, "mathIndex": 70.0, "agenticIndex": 58.0, "reasoningIndex": 68.0,
+    "mmluPro": 75.0, "gpqa": 52.0, "hle": 8.4, "contextWindow": 128000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "DeepSeek Chat"
+  },
+  "z-ai/glm-5.2": {
+    "codingIndex": 82.4, "mathIndex": 79.0, "agenticIndex": 76.1, "reasoningIndex": 78.9,
+    "mmluPro": 84.2, "gpqa": 61.8, "hle": 13.5, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GLM 5.2"
+  },
+  "zai-glm-4.7": {
+    "codingIndex": 75.0, "mathIndex": 73.5, "agenticIndex": 68.4, "reasoningIndex": 72.0,
+    "mmluPro": 78.5, "gpqa": 55.2, "hle": 10.8, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GLM 4.7"
+  },
+  "@cf/zai-org/glm-4.7-flash": {
+    "codingIndex": 70.0, "mathIndex": 69.0, "agenticIndex": 62.0, "reasoningIndex": 66.0,
+    "mmluPro": 73.0, "gpqa": 50.0, "hle": 9.0, "contextWindow": 128000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GLM 4.7 Flash"
+  },
+  "moonshotai/kimi-k2.6": {
+    "codingIndex": 79.5, "mathIndex": 77.0, "agenticIndex": 73.2, "reasoningIndex": 76.4,
+    "mmluPro": 82.0, "gpqa": 59.0, "hle": 12.4, "contextWindow": 262000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Kimi K2.6"
+  },
+  "@cf/moonshotai/kimi-k2.6": {
+    "codingIndex": 79.5, "mathIndex": 77.0, "agenticIndex": 73.2, "reasoningIndex": 76.4,
+    "mmluPro": 82.0, "gpqa": 59.0, "hle": 12.4, "contextWindow": 262000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Kimi K2.6"
+  },
+  "@cf/moonshotai/kimi-k2.7-code": {
+    "codingIndex": 81.0, "mathIndex": 76.5, "agenticIndex": 74.0, "reasoningIndex": 75.5,
+    "mmluPro": 81.5, "gpqa": 58.0, "hle": 12.0, "contextWindow": 262000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Kimi K2.7 Code"
+  },
+  "minimaxai/minimax-m2.7": {
+    "codingIndex": 77.5, "mathIndex": 75.0, "agenticIndex": 70.0, "reasoningIndex": 74.5,
+    "mmluPro": 80.5, "gpqa": 56.0, "hle": 11.5, "contextWindow": 200000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "MiniMax M2.7"
+  },
+  "minimaxai/minimax-m3": {
+    "codingIndex": 78.0, "mathIndex": 76.0, "agenticIndex": 71.0, "reasoningIndex": 75.0,
+    "mmluPro": 81.0, "gpqa": 57.0, "hle": 12.0, "contextWindow": 1000000,
+    "supportsReasoning": true, "supportsVision": true,
+    "lastUpdated": "2026-07-20", "originalModel": "MiniMax M3"
+  },
+  "MiniMax-M2.7": {
+    "codingIndex": 77.5, "mathIndex": 75.0, "agenticIndex": 70.0, "reasoningIndex": 74.5,
+    "mmluPro": 80.5, "gpqa": 56.0, "hle": 11.5, "contextWindow": 192000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "MiniMax M2.7"
+  },
+  "stepfun-ai/step-3.7-flash": {
+    "codingIndex": 73.5, "mathIndex": 74.0, "agenticIndex": 68.0, "reasoningIndex": 72.0,
+    "mmluPro": 78.0, "gpqa": 53.0, "hle": 10.5, "contextWindow": 256000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Step 3.7 Flash"
+  },
+  "openai/gpt-oss-120b": {
+    "codingIndex": 62.0, "mathIndex": 64.0, "agenticIndex": 58.0, "reasoningIndex": 63.0,
+    "mmluPro": 70.0, "gpqa": 52.0, "hle": 9.0, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GPT OSS 120B"
+  },
+  "gpt-oss-120b": {
+    "codingIndex": 62.0, "mathIndex": 64.0, "agenticIndex": 58.0, "reasoningIndex": 63.0,
+    "mmluPro": 70.0, "gpqa": 52.0, "hle": 9.0, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GPT OSS 120B"
+  },
+  "openai/gpt-oss-20b": {
+    "codingIndex": 50.0, "mathIndex": 52.0, "agenticIndex": 46.0, "reasoningIndex": 51.0,
+    "mmluPro": 58.0, "gpqa": 42.0, "hle": 6.5, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GPT OSS 20B"
+  },
+  "gpt-oss-20b": {
+    "codingIndex": 50.0, "mathIndex": 52.0, "agenticIndex": 46.0, "reasoningIndex": 51.0,
+    "mmluPro": 58.0, "gpqa": 42.0, "hle": 6.5, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "GPT OSS 20B"
+  },
+  "meta/llama-4-maverick-17b-128e-instruct": {
+    "codingIndex": 73.5, "mathIndex": 70.0, "agenticIndex": 65.0, "reasoningIndex": 71.0,
+    "mmluPro": 78.0, "gpqa": 53.0, "hle": 10.0, "contextWindow": 1000000,
+    "supportsReasoning": false, "supportsVision": true,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 4 Maverick"
+  },
+  "meta-llama/llama-4-scout-17b-16e-instruct": {
+    "codingIndex": 28.0, "mathIndex": 25.0, "agenticIndex": 22.0, "reasoningIndex": 26.0,
+    "mmluPro": 30.0, "gpqa": 18.0, "hle": 2.0, "contextWindow": 10000000,
+    "supportsReasoning": false, "supportsVision": true,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 4 Scout"
+  },
+  "@cf/meta/llama-4-scout-17b-16e-instruct": {
+    "codingIndex": 28.0, "mathIndex": 25.0, "agenticIndex": 22.0, "reasoningIndex": 26.0,
+    "mmluPro": 30.0, "gpqa": 18.0, "hle": 2.0, "contextWindow": 10000000,
+    "supportsReasoning": false, "supportsVision": true,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 4 Scout"
+  },
+  "meta/llama-3.3-70b-versatile": {
+    "codingIndex": 22.0, "mathIndex": 24.0, "agenticIndex": 18.0, "reasoningIndex": 23.0,
+    "mmluPro": 28.0, "gpqa": 16.0, "hle": 1.5, "contextWindow": 131072,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 3.3 70B Versatile"
+  },
+  "llama-3.3-70b-versatile": {
+    "codingIndex": 22.0, "mathIndex": 24.0, "agenticIndex": 18.0, "reasoningIndex": 23.0,
+    "mmluPro": 28.0, "gpqa": 16.0, "hle": 1.5, "contextWindow": 131072,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 3.3 70B Versatile"
+  },
+  "Meta-Llama-3.3-70B-Instruct": {
+    "codingIndex": 22.0, "mathIndex": 24.0, "agenticIndex": 18.0, "reasoningIndex": 23.0,
+    "mmluPro": 28.0, "gpqa": 16.0, "hle": 1.5, "contextWindow": 128000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 3.3 70B Instruct"
+  },
+  "llama-3.1-8b-instant": {
+    "codingIndex": 18.0, "mathIndex": 19.0, "agenticIndex": 14.0, "reasoningIndex": 17.0,
+    "mmluPro": 22.0, "gpqa": 12.0, "hle": 1.0, "contextWindow": 131072,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Llama 3.1 8B Instant"
+  },
+  "qwen/qwen3.6-27b": {
+    "codingIndex": 76.5, "mathIndex": 74.0, "agenticIndex": 70.0, "reasoningIndex": 73.0,
+    "mmluPro": 79.0, "gpqa": 55.0, "hle": 11.0, "contextWindow": 131072,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Qwen3.6 27B"
+  },
+  "qwen/qwen2.5-coder-32b-instruct": {
+    "codingIndex": 48.0, "mathIndex": 45.0, "agenticIndex": 42.0, "reasoningIndex": 44.0,
+    "mmluPro": 52.0, "gpqa": 32.0, "hle": 4.0, "contextWindow": 128000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Qwen2.5 Coder 32B"
+  },
+  "qwen/qwen3-30b-a3b-fp8": {
+    "codingIndex": 65.0, "mathIndex": 64.0, "agenticIndex": 58.0, "reasoningIndex": 63.0,
+    "mmluPro": 70.0, "gpqa": 48.0, "hle": 8.0, "contextWindow": 131072,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Qwen3 30B A3B"
+  },
+  "mistralai/mistral-medium-3.5-128b": {
+    "codingIndex": 76.5, "mathIndex": 74.0, "agenticIndex": 70.5, "reasoningIndex": 73.0,
+    "mmluPro": 79.0, "gpqa": 56.0, "hle": 11.2, "contextWindow": 256000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Mistral Medium 3.5"
+  },
+  "mistralai/mistral-small-4-119b-2603": {
+    "codingIndex": 60.0, "mathIndex": 58.0, "agenticIndex": 55.0, "reasoningIndex": 57.0,
+    "mmluPro": 65.0, "gpqa": 44.0, "hle": 7.0, "contextWindow": 256000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Mistral Small 4"
+  },
+  "mistralai/mistral-large-3-675b-instruct-2512": {
+    "codingIndex": 58.0, "mathIndex": 56.0, "agenticIndex": 52.0, "reasoningIndex": 55.0,
+    "mmluPro": 63.0, "gpqa": 42.0, "hle": 6.5, "contextWindow": 256000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Mistral Large 675B"
+  },
+  "mistralai/ministral-14b-instruct-2512": {
+    "codingIndex": 34.0, "mathIndex": 32.0, "agenticIndex": 28.0, "reasoningIndex": 31.0,
+    "mmluPro": 38.0, "gpqa": 22.0, "hle": 3.0, "contextWindow": 32000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Ministral 14B"
+  },
+  "google/gemma-4-31b-it": {
+    "codingIndex": 52.0, "mathIndex": 50.0, "agenticIndex": 46.0, "reasoningIndex": 49.0,
+    "mmluPro": 58.0, "gpqa": 38.0, "hle": 5.5, "contextWindow": 256000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Gemma 4 31B"
+  },
+  "gemma-4-31b": {
+    "codingIndex": 52.0, "mathIndex": 50.0, "agenticIndex": 46.0, "reasoningIndex": 49.0,
+    "mmluPro": 58.0, "gpqa": 38.0, "hle": 5.5, "contextWindow": 128000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Gemma 4 31B"
+  },
+  "@cf/google/gemma-4-26b-a4b-it": {
+    "codingIndex": 48.0, "mathIndex": 46.0, "agenticIndex": 42.0, "reasoningIndex": 45.0,
+    "mmluPro": 54.0, "gpqa": 35.0, "hle": 5.0, "contextWindow": 128000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Gemma 4 26B"
+  },
+  "nvidia/nemotron-3-ultra-550b-a55b": {
+    "codingIndex": 71.0, "mathIndex": 70.0, "agenticIndex": 65.0, "reasoningIndex": 69.0,
+    "mmluPro": 75.0, "gpqa": 52.0, "hle": 9.5, "contextWindow": 1000000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Nemotron 3 Ultra"
+  },
+  "nvidia/nemotron-3-super-120b-a12b": {
+    "codingIndex": 60.0, "mathIndex": 59.0, "agenticIndex": 54.0, "reasoningIndex": 58.0,
+    "mmluPro": 65.0, "gpqa": 44.0, "hle": 7.0, "contextWindow": 128000,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Nemotron 3 Super"
+  },
+  "nvidia/nemotron-3-nano-30b-a3b": {
+    "codingIndex": 38.0, "mathIndex": 36.0, "agenticIndex": 32.0, "reasoningIndex": 35.0,
+    "mmluPro": 42.0, "gpqa": 26.0, "hle": 3.5, "contextWindow": 1000000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Nemotron 3 Nano"
+  },
+  "bytedance/seed-oss-36b-instruct": {
+    "codingIndex": 56.0, "mathIndex": 54.0, "agenticIndex": 50.0, "reasoningIndex": 53.0,
+    "mmluPro": 60.0, "gpqa": 40.0, "hle": 6.0, "contextWindow": 32000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Seed OSS 36B"
+  },
+  "stockmark/stockmark-2-100b-instruct": {
+    "codingIndex": 36.0, "mathIndex": 34.0, "agenticIndex": 30.0, "reasoningIndex": 33.0,
+    "mmluPro": 40.0, "gpqa": 24.0, "hle": 3.0, "contextWindow": 32000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Stockmark 100B"
+  },
+  "codestral-latest": {
+    "codingIndex": 42.0, "mathIndex": 40.0, "agenticIndex": 38.0, "reasoningIndex": 39.0,
+    "mmluPro": 48.0, "gpqa": 30.0, "hle": 4.0, "contextWindow": 32000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Codestral Latest"
+  },
+  "qwen/qwq-32b": {
+    "codingIndex": 58.0, "mathIndex": 82.0, "agenticIndex": 50.0, "reasoningIndex": 76.0,
+    "mmluPro": 75.0, "gpqa": 64.0, "hle": 11.0, "contextWindow": 131072,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "QwQ 32B"
+  },
+  "poolside/laguna-m.1:free": {
+    "codingIndex": 72.0, "mathIndex": 68.0, "agenticIndex": 65.0, "reasoningIndex": 69.0,
+    "mmluPro": 74.0, "gpqa": 50.0, "hle": 9.5, "contextWindow": 262000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Poolside Laguna M.1"
+  },
+  "poolside/laguna-xs.2:free": {
+    "codingIndex": 68.0, "mathIndex": 65.0, "agenticIndex": 62.0, "reasoningIndex": 65.0,
+    "mmluPro": 70.0, "gpqa": 47.0, "hle": 8.5, "contextWindow": 262000,
+    "supportsReasoning": false, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Poolside Laguna XS.2"
+  },
+  "@cf/qwen/qwen3-30b-a3b-fp8": {
+    "codingIndex": 65.0, "mathIndex": 64.0, "agenticIndex": 58.0, "reasoningIndex": 63.0,
+    "mmluPro": 70.0, "gpqa": 48.0, "hle": 8.0, "contextWindow": 131072,
+    "supportsReasoning": true, "supportsVision": false,
+    "lastUpdated": "2026-07-20", "originalModel": "Qwen3 30B A3B"
+  }
+}