free-coding-models 0.5.83 → 0.5.85

package/README.md

@@ -192,6 +192,7 @@ A large catalog of coding models from 20+ active providers, ranked by practical
 | 21 | [SiliconFlow](https://cloud.siliconflow.cn/account/ak) | 3 | A → B+ | 3 models @ $0, 1000 RPM (phone SMS required) | `SILICONFLOW_API_KEY` |
 | 22 | [Requesty](https://app.requesty.ai/api-keys) | 12 | S+ → C | 200 req/day free, no card (4x OpenRouter) | `REQUESTY_API_KEY` |
 | 23 | [OrcaRouter](https://www.orcarouter.ai) | 3 | S+ → A+ | Free Hacker tier · 3 $-0 models, zero token markup | `ORCAROUTER_API_KEY` |
+| 24 | [Vercel AI Gateway](https://vercel.com/ai-gateway) | 4 | S+ → B+ | $5 gateway credits every 30 days, no card + explicit $0 models | `VERCEL_AI_GATEWAY_API_KEY` |
 
 > **Pollinations Pollen, c'est quoi ?** Pollinations ne facture pas en $ mais en **Pollen** (≈ $1). Pas besoin d'acheter : tu recois du Pollen **gratuit chaque jour** selon ton tier seed/flower/nectar via quests (star le repo `pollinations/pollinations` sur GitHub, aider sur un issue). Si ton solde affiche `0.0000` sur `enter.pollinations.ai`, tout `MISS` (prompt non cache) retourne `402 PAYMENT_REQUIRED` ("Insufficient balance ~0.0001 pollen") et le ping affichera `402`. Un `HIT` Cloudflare cache peut encore passer en `200` meme avec 0 Pollen mais ce n'est pas fiable. Verifie ton solde sur `enter.pollinations.ai` -> API Keys, complete une quest et attends le grant quotidien a minuit UTC. **Note : integration Pollinations encore experimentale et non confirmee comme stable, le systeme Pollen peut changer.**
 > **SiliconFlow & Requesty** : integrations verifiees le 2026-08-24 via `/v1/models` live et docs officielles, free tiers recurrents confirmes sans Pollen.
@@ -569,7 +570,7 @@ On first start the daemon auto-creates a `fast-coding` set from your configured
 | 🟠 **Auth error** | 401/403 — your API key is wrong for this model |
 | ⚪ **Deprecated** | Removed from the catalog — will be replaced |
 
-**3. Failover order** — models are tried in priority order; a `Recovering` / `Down` / `Auth error` model is skipped and the request goes to the next healthy one. If *all* fail, you get a `503` with a `models_tried` list in the body for debugging.
+**3. Failover order (family preserving)** — models are tried in priority order; a `Recovering` / `Down` / `Auth error` model is skipped and the request goes to the next healthy one. When a model fails, the router first looks for the **same model family on another provider** (e.g. `nvidiaNim/deepseek-v4-pro` down → `sambanova/DeepSeek-V3.1` next), so your coding assistant keeps the same model behaviour mid-conversation instead of silently switching to a different family. If no same-family alternative is healthy, it falls back to the plain priority order. Families covered: Claude, DeepSeek, Gemini, GPT, Nemotron, Llama, MiniMax, Qwen, Kimi, GLM, Mistral, OpenAI o-series. The toggle lives on the Router Dashboard's set panel (**Family failover: on/off**, on by default, stored per set as `familyFailover`), and family hops are tagged `family` in the request log of both the TUI and Web dashboard. If *all* fail, you get a `503` with a `models_tried` list in the body for debugging.
 
 **4. Auto-heal (on by default)** — at daemon start, any model in `Auth error` or `Deprecated` is swapped for a working alternative (same provider first, then cross-provider). The first time you manually add/remove/reorder a model, auto-heal switches off and your choices are preserved — so a new user with a half-broken key set lands on a usable default set by the time the dashboard renders.
 
@@ -985,6 +986,7 @@ Curated resources kept **outside** the active CLI catalog — IDE extensions, co
 
 | Provider | Free limits | Notable models |
 |----------|-------------|----------------|
+| [Vercel AI Gateway](https://vercel.com/ai-gateway) | $5 credits every 30 days (no card) + explicit $0 models | MiniMax M3 (1M ctx), Laguna S 2.1, MiniMax M2.7, Ling 3.0 Flash Fin |
 | [OrcaRouter](https://www.orcarouter.ai) | Free Hacker tier, zero token markup, 3 API keys | DeepSeek V4 Flash (Free), Tencent Hy3 (Free), Qwen3.8 27B (Free) |
 | [OpenRouter](https://openrouter.ai/keys) | 50 req/day, 1K/day with $10 purchase | Qwen3-Coder, Tencent HY3, Laguna, Gemma 4 |
 | [Google AI Studio](https://aistudio.google.com/apikey) | Varies by Gemini model and region | Gemini 3.8 Flash, Gemini 2.5 Pro |

package/changelog/v0.5.84.md

@@ -0,0 +1,32 @@
+# Changelog v0.5.84 - 2026-09-05
+
+### Changed
+
+- Full catalog audit of all 23 providers against official docs and live model APIs (2026-09-05): 28 removals, 42 additions, 37 config fixes, 147 models confirmed unchanged. Catalog goes from 208 to 225 models.
+
+Provider by provider:
+
+- **Kilo**: the free gateway exploded from 2 to 19 models. Added MiniMax M3 (80.5% SWE-bench Verified, 1M ctx), Thinking Machines Inkling + Inkling Small (1M ctx), StepFun Step 3.7 Flash, Poolside Laguna S 2.1, NVIDIA Nemotron 3 Ultra + 3 Super + 3.5 Lightning, Cohere North Mini Code, Dots3-Note Preview, MiniMax M2.7 and the OpenRouter free router alias. `kilo-auto/small` is now paid and was removed.
+- **NVIDIA NIM**: removed models that are gone or deprecated (gpt-oss-120b API deprecated 2026-09-02, kimi-k2.6 404, nemotron-nano-3, llama-3.2-11b-vision without endpoint). Added DeepSeek V4 Pro (1M ctx) and Nemotron 3.5 Lightning 30B. kimi-k3 and nemotron-3-super now serve 1M contexts. MiniMax M3 is kept with a deprecation notice (NVIDIA shutdown 2026-09-08).
+- **Scaleway**: applied the July 2026 EOL wave. Removed Devstral 2, Gemma 3 27B and Holo2 (EOL already passed). Qwen3 Coder 30B and Pixtral 12B stay listed with an EOL 2026-10-01 warning since they still answer. The service now bills per token with the first 1M tokens free.
+- **Routeway**: the free catalog was gutted from 13 to 4 models. Kept DeepSeek V4 Flash, MiniMax M2.7 and Muse Glimmer 30B (contexts slashed to 42k/42k/131k) and added Kimi K2.6 as the only new free model. 10 free variants are now paid-only.
+- **Cloudflare Workers AI**: kimi-k2.6, kimi-k2.7-code and glm-5.2 moved behind Workers Paid billing and were removed from the free list. Fixed 7 inflated context windows (docs say 24k to 256k, not 128k). Added qwen3.8-27b (262k ctx) as the only new free coder-usable model.
+- **Qwen / Model Studio**: added the new Qwen3.8 family (qwen3.8-max, qwen3.8-flash, qwen3.8-2.4t-a95b, qwen3.8-27b), each with 1M context and a 1M-token international free quota. All 23 existing models verified unchanged.
+- **Z.ai**: added the new coding-plan flagship GLM-5.3 (1M ctx, released 2026-08-18). Filled in official SWE-bench Verified scores for GLM-4.7 (73.8%) and GLM-4.6 (68.0%).
+- **Mistral**: the three Ministral 3 models now serve 256k context (was 128k). Added Z.ai GLM 5.2 (1M ctx, public preview on La Plateforme).
+- **Groq**: added qwen/qwen3.8-27b preview. Fixed gpt-oss-20b SWE-bench score to the official 60.7%. None of the catalog models are on the deprecation list.
+- **Cerebras**: gemma-4-31b was pulled from free public endpoints on 2026-09-03 and is replaced by qwen-3.8-27b (64k free ctx).
+- **Pollinations**: added 3 anonymous-free models: glm-5.3 (1M ctx), Kimi K2.6 (80.2%) and MiniMax M3 (80.5%).
+- **SiliconFlow**: DeepSeek-R1-Distill-Qwen-7B was delisted. Added DeepSeek-R1-0528-Qwen3-8B, Qwen3.5-4B, Qwen2.5-7B-Instruct and GLM-4-9B-0414. Both surviving models actually serve 131k context, not 32k.
+- **OpenRouter**: fixed 7 wrong context windows against the live API (GLM-5.2 is 256k, not 1M; Inkling models are 1M; MiniMax M2.7 is 192k). Added ling-3.0-flash-sante:free.
+- **Ollama Cloud**: added glm-5.3 and glm-5.3-flash. Fixed context windows for glm-5.2 (976k), minimax-m3 (512k) and gemma4:31b (256k).
+- **Novita**: the empty free catalog is restored with 2 time-limited $0/M models: Ling 3.0 Flash Fin and Ling 3.0 Flash Sante (256k ctx, coding + function calling).
+- **SambaNova**: MiniMax-M3 context is now 1M tokens (was 196k).
+- **Codestral**: context fixed to 256k. Still free via Mistral's Experiment tier (~1B tokens/month).
+- **Requesty**: fixed the model id typo `nvidia/nemotron-nano-3-30b-a3b` to the real `nvidia/nemotron-3-nano-30b-a3b`.
+- **OVHcloud**: added 2 free Qwen3Guard moderation models (beta) and fixed Qwen2.5-VL-72B context to 32k. The chat LLMs are now paid per Mtoken but stay listed for now since they still answer.
+- **googleai, llm7, orcarouter**: verified live and unchanged, nothing to update.
+
+### Fixed
+
+- Tests: the t8 family-failover tests now use `openai/gpt-oss-20b` as NVIDIA primary because gpt-oss-120b was removed from the NVIDIA catalog (deprecated 2026-09-02). The NVIDIA test-model discovery expectations match the new catalog head.

package/changelog/v0.5.85.md

@@ -0,0 +1,21 @@
+# Changelog v0.5.85 - 2026-09-05
+
+### Added
+
+- New provider: **Vercel AI Gateway** (24th provider in the catalog). This is Vercel's official OpenAI-compatible gateway: one endpoint at `https://ai-gateway.vercel.sh/v1/chat/completions`, list prices with zero markup, and it works from your local machine with nothing deployed to Vercel.
+  - **Free tier**: every Vercel account gets **$5 of gateway credits every 30 days**, no credit card required. On top of that, the gateway exposes a handful of genuinely $0 models (input AND output priced at $0), and those are the ones we list.
+  - **4 models added**, all verified live against `GET /v1/models` (373 models scanned, 2026-09-05):
+    - `minimax/minimax-m3-free` - MiniMax M3 (Free), S+ 78.4%, 1M context, vision + tool use
+    - `minimax/minimax-m2.7-free` - MiniMax M2.7 (Free), S+ 78.0%, 196k context, software-engineering focused
+    - `poolside/laguna-s-2.1-free` - Laguna S 2.1 (Free), S+ tier, 256k context, dedicated agentic coding model
+    - `inclusionai/ling-3.0-flash-fin-free` - Ling 3.0 Flash Fin (Free), B+ tier, 256k context, finance-tuned 124B MoE that retains coding and math
+  - **Quota marked `limited`** on purpose: the monthly credit only covers a subset of the 373-model catalog, and per Vercel's FAQ purchasing credits once permanently moves the account to the paid tier, which kills the recurring free credit. The Settings panel explains this.
+  - **Full cross-surface wiring**: TUI Settings panel entry with its own theme color, key test endpoint (`/v1/models`), web dashboard catalog, OpenCode + Kilo provider config generation, and monthly reset cadence for quota display. Env var: `VERCEL_AI_GATEWAY_API_KEY` (also accepts `AI_GATEWAY_API_KEY`).
+  - Get a key: Vercel dashboard, AI Gateway → API Keys.
+- New unit test covering the `vercel-gateway` source entry structure (model IDs must be explicit `-free` $0 models). Suite now at 880 tests.
+
+### Notes
+
+- MiniMax M3/M2.7 scores mirror the measured entries already in the catalog (ollama-cloud, benchmarked 2026-07-27); the underlying models are identical.
+- Laguna S 2.1 has no published SWE-bench score; its S+ tier follows the family precedent (Laguna XS 2.1 ships S+ 70.9% via NVIDIA). models.dev flags the model as deprecated as of 2026-09-05 (the Zen free promo ended), but Vercel still serves the $0 variant live today, so it stays listed with a re-verify note for the next audit.
+- Catalog totals: 24 providers, 229 models.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "free-coding-models",
-  "version": "0.5.83",
+  "version": "0.5.85",
   "description": "Find the fastest coding LLM models in seconds — ping free models from multiple providers, pick the best one for OpenCode, Cursor, or any AI coding assistant.",
   "keywords": [
     "nvidia",
@@ -58,7 +58,7 @@
   ],
   "scripts": {
     "start": "node bin/free-coding-models.js",
-    "test": "node --test test/test.js test/fcm-agent-core.test.js test/patch-openclaw.test.js test/provider-metadata.test.js test/config-permission-hint.test.js test/probe-cache.test.js test/passive-quota.test.js test/runtime-telemetry.test.js test/extended-benchmarks.test.js test/models-dev.test.js test/model-merger.test.js test/config-dir.test.js test/cache-dir.test.js",
+    "test": "node --test test/test.js test/fcm-agent-core.test.js test/patch-openclaw.test.js test/provider-metadata.test.js test/config-permission-hint.test.js test/probe-cache.test.js test/passive-quota.test.js test/runtime-telemetry.test.js test/extended-benchmarks.test.js test/models-dev.test.js test/model-merger.test.js test/config-dir.test.js test/cache-dir.test.js test/model-family.test.js",
     "test:extended-benchmarks": "node --test test/extended-benchmarks.test.js",
     "test:models-dev": "node --test test/models-dev.test.js",
     "update:benchmarks": "node scripts/update-benchmarks.mjs",