free-coding-models 0.5.97 → 0.5.99

package/README.md

@@ -6,7 +6,7 @@
 
 <p align="center">
   <strong>Find the fastest free coding model in seconds.</strong><br>
-  Live latency, stability and verdicts for 271 models from 25 free AI providers, then install the one you pick straight into your favorite coding tool.<br><br>
+  Live latency, stability and verdicts for 252 models from 24 free AI providers, then install the one you pick straight into your favorite coding tool.<br><br>
   <strong>Works with:</strong> OpenCode CLI / Desktop / WebUI, OpenClaw, Crush, Goose, Aider, Kilo CLI, Qwen Code, OpenHands, Amp, Hermes, Continue, Cline, Xcode, Pi, ZCode, ForgeCode, Copilot, jcode, Caveman Code and more.
 </p>
 
@@ -24,7 +24,7 @@ free-coding-models
 ```
 
 <p align="center">
-  <sub>Then create a free account on one of the <a href="#-providers">25 providers</a> to grab an API key.</sub>
+  <sub>Then create a free account on one of the <a href="#-providers">24 providers</a> to grab an API key.</sub>
 </p>
 
 <p align="center">
@@ -35,7 +35,7 @@ free-coding-models
 
 ## 💡 Why this tool?
 
-There is a large catalog of free and free-limited coding models (**25 providers / 271 live models**, generated from [`sources.js`](./sources.js)). Which one is fastest *right now*? Which one is actually stable, versus just lucky on the last ping?
+There is a large catalog of free and free-limited coding models (**24 providers / 252 live models**, generated from [`sources.js`](./sources.js)). Which one is fastest *right now*? Which one is actually stable, versus just lucky on the last ping?
 
 `free-coding-models` (FCM) answers that by pinging every model in parallel, showing live latency, and computing a **live Stability Score (0-100)** combining p95 latency, jitter, spike rate and uptime. Average latency alone is misleading: a model that randomly spikes to 6 seconds is not reliable.
 
@@ -87,18 +87,18 @@ free-coding-models --fiable              # print the single most reliable model
 
 ## 🟢 Providers
 
-**25 active providers / 271 live models**, sorted by live model count. Top 8:
+**24 active providers / 252 live models**, sorted by live model count. Top 8:
 
 | Provider | Models | Best tier | Env var |
 |----------|--------|-----------|---------|
-| [onomeo](https://onomeo.com/dashboard) | 33 | S+ | `ONOMEO_API_KEY` |
-| [Alibaba DashScope](https://modelstudio.console.alibabacloud.com) | 29 | S+ | `DASHSCOPE_API_KEY` |
-| [Pollinations AI](https://enter.pollinations.ai) | 21 | S+ | `POLLINATIONS_API_KEY` |
-| [OpenRouter](https://openrouter.ai/keys) | 18 | S+ | `OPENROUTER_API_KEY` |
-| [Kilo](https://kilo.ai) | 17 | S+ | `KILO_API_KEY` |
-| [Ollama Cloud](https://ollama.com/settings/keys) | 17 | S+ | `OLLAMA_API_KEY` |
-| [NVIDIA NIM](https://build.nvidia.com) | 16 | S+ | `NVIDIA_API_KEY` |
-| [OVHcloud AI](https://endpoints.ai.cloud.ovh.net) | 15 | S+ | `OVH_AI_ENDPOINTS_ACCESS_TOKEN` |
+| [Pollinations AI](https://enter.pollinations.ai) | 25 | S+ | `POLLINATIONS_API_KEY` |
+| [Alibaba DashScope](https://modelstudio.console.alibabacloud.com) | 22 | S+ | `DASHSCOPE_API_KEY` |
+| [Ollama Cloud](https://ollama.com/settings/keys) | 19 | S+ | `OLLAMA_API_KEY` |
+| [Kilo](https://kilo.ai) | 16 | S+ | `KILO_API_KEY` |
+| [OpenRouter](https://openrouter.ai/keys) | 16 | S+ | `OPENROUTER_API_KEY` |
+| [onomeo](https://onomeo.com/dashboard) | 15 | S+ | `ONOMEO_API_KEY` |
+| [NVIDIA NIM](https://build.nvidia.com) | 14 | S+ | `NVIDIA_API_KEY` |
+| [OVHcloud AI](https://endpoints.ai.cloud.ovh.net) | 14 | S+ | `OVH_AI_ENDPOINTS_ACCESS_TOKEN` |
 
 > 🧾 **What "free" means here:** free is a property of the *(provider, model)* pair, never of the provider as a whole. A row is listed only when that exact model id costs $0 to call through that provider (permanent free tier, `:free` variant, or free plan), verified live at audit time. The same open-weights model can be free on one host and paid on another - paid siblings are deliberately excluded. Full breakdown and badge legend: [`docs/providers.md`](./docs/providers.md).
 

package/changelog/v0.5.98.md

@@ -0,0 +1,39 @@
+# Changelog v0.5.98 - 2026-10-10
+
+Full 25-provider audit against live provider APIs, official catalogs and deprecation notices. The catalog goes from 271 to 253 models: 46 dead or now-paid models removed, 28 fresh free coding models added, 9 metadata fixes.
+
+### Added
+
+- **Step 5 Preview lands on 4 gateways**: Kilo (`stepfun/step-5-preview-free`, S, 1M ctx), OpenCode Zen (`step-5-preview-free`), llm7 and Pollinations - the StepFun flagship preview is this week's big free-model rollout.
+- **Ling 3.1 Flash (560B-A25B MoE) replaces the retired Ling 3.0 :free slots** on OpenRouter ($0/$0), Kilo, OpenCode Zen, Requesty, Novita and Vercel AI Gateway (`ling-3.1-flash-free`).
+- **Glyph Cluster, the new free stealth coding model**, takes the stealth slot on Kilo and Vercel AI Gateway (Pixel Canary rotated out).
+- **llm7 rebuilds its free tier**: + GLM-5.3-Flash (400k ctx, turbo tier), + Kimi K2.7 Code, + Kimi K2.6, + GPT-OSS 20B, + Nemotron 3 Nano 30B (1M ctx) - the catalog grows from 4 to 8 models.
+- **Novita reshuffles around the Qwen3.8 generation**: + GLM-5.3-Flash (1M ctx), + Qwen3.8 Flash (1M ctx), + Ling 3.0 Flash, + GLM-4.7-Flash, + MiniMax M2.7.
+- **Pollinations**: + OpenAI GPT-6.1 Sol (S+, 1M ctx), + xAI Grok 4.7 (S, 500k ctx), + StepFun Step 5 Preview (A+, 1M ctx).
+- **onomeo**: + Gemini 3.6 Flash (S+, 1M ctx) and + Hunyuan 4 Preview (1M ctx), replacing the removed Hunyuan 3 slot.
+- **Mistral La Plateforme**: + Mistral Vibe CLI (new free coding model) and + Leanstral 1.5.1 (successor of Leanstral 1.5).
+- **Ollama Cloud**: + Mistral Large 4 (1.05T MoE, 1M ctx, ~Oct 4 release).
+- **OpenRouter**: + Apodex 1.1 Mini (`:free`, 262k ctx).
+
+### Changed
+
+- **qwen (Alibaba DashScope): 5 models retired exactly today, 2026-10-10**, as predicted by the in-file deprecation notices: qwen3.6-max-preview, qwen3-coder-plus, qwen3-coder-next, qwen3-coder-480b-a35b-instruct and qwen3-coder-30b-a3b-instruct (notices 1949/1950), plus qwen3-vl-flash. Also dropped the phantom qwen3.8-flash-next (404 detail page, never existed in the product index). 29 -> 22 models.
+- **onomeo: 16 of 33 models removed** following the provider's own removed list, and mirrored ids renamed to first-party ones (`openai/gpt-oss-20b` -> `gpt-oss-20b`, `z-ai/glm-5.3-flash-free` -> `glm-5.3-flash`, `meta/muse-glimmer-30b` -> `muse-glimmer-30b`). 5 remaining models are flagged as upstream-paused.
+- **nvidia: the 3 models re-added on 2026-09-29 are ghost catalog entries again** (page 404 + absent from the hosted list): Kimi K2.6, Nemotron Nano 3 30B, Codestral 22B removed; 16 -> 13 models.
+- **sambanova: MiniMax M2.7 and M3 went paid-only** - the official Free Tier table now covers exactly 5 models, the MiniMax pair needs a linked payment method (Developer Tier). 7 -> 5 models.
+- **kilo: hard free-tier rotation** - qwen3.8-27b:free, step-3.7:free, ling-3.0-flash-sante:free and space-bunny-alpha gone; note: paid Poolside Laguna variants retire Oct 31, the free ones may follow.
+- **openrouter**: qwen3.8-27b:free deleted from the live catalog, ling-3.0-flash-sante went paid, stealth/space-bunny-alpha rotated out.
+- **novita**: the Qwen3.5/3.6 Plus tier was killed in favor of Qwen3.8 (2 removals).
+- **ovhcloud**: Qwen3-Coder-30B is dead again (model_not_found on live probe). Warning from the audit: OVH now bills most text models per token - only a handful stay tagged Free on the catalog page.
+- **Moderation classifiers cleaned out of the coding catalog**: Nemotron 3.5 Content Safety removed from OpenRouter, Kilo and Requesty (guardrail model, not coding-capable).
+- README counts and the Top 8 provider table refreshed; `docs/providers.md` regenerated from `sources.js`.
+
+### Fixed
+
+- qwen omni context windows were wrong: qwen3.5-omni-plus 32k -> 256k, qwen3.8-omni-flash 32k -> 1M (official model pages).
+- onomeo: gpt-oss-20b and muse-glimmer-30b ctx 128k -> 131k, minimax-m2.7 ctx 200k -> 205k (live catalog).
+- kilo: openrouter/free ctx dropped (the live gateway no longer publishes a context length for the router).
+
+### Deliberately NOT added (audit transparency)
+
+15 researcher-reported additions were skipped because they reversed verdicts already recorded in `sources.js` from earlier audits, with no stronger evidence: the 6 "free again" SiliconFlow small models (API needs a paid-check key to verify), 4 Pollinations models flagged "paid Pollen" on 2026-09-26/28 (the issue #190 trap: they return HTTP 200 with an embedded credits error, and this audit did not sniff completion bodies), QwQ 32B + R1-Distill on Cloudflare (supersession decision re-confirmed 2026-09-29), the twice-removed Mistral 7B v0.3 / Nemo 12B on OVHcloud, and gemini-3.1-pro-preview on Google AI (excluded twice as paid-only). All are documented in `sources.js` comments for the next audit to re-verify with live keys.

package/changelog/v0.5.99.md

@@ -0,0 +1,17 @@
+# Changelog v0.5.99 - 2026-10-17
+
+### Changed
+
+- 📡 **Full 24-provider audit** (weekly): 253 → 252 models. 29 models removed, 28 added, 18 config fixes, all verified against live provider APIs and official docs on 2026-10-17. `audit_state.json` refreshed for all live providers (the stale `github-models` entry was dropped along with `zai`).
+- 🚀 **Reflection Beam 1.0 everywhere**: the new 501B MoE agentic coder (SWE-bench Verified 80.9-81.6%) landed free on 4 providers the same week - Kilo, OpenCode Zen, Vercel AI Gateway and Ollama Cloud. It is now the top free coder in the catalog.
+- ➕ **New free models**: NVIDIA NIM adds Nemotron 3.5 Super VL (71.07% SWE, multimodal) and StepFun Step 5 Preview (600B MoE, 1M ctx); XingChen Xing4.0 29B (75.0% SWE) joins OpenRouter and Kilo; MiniMax M3 (80.5% SWE, 1M ctx) on SambaNova was proposed but NOT re-added - a `/v1/models` listing is not free-tier evidence and the official Free Tier table still excludes it (the 2026-10-10 verdict stands, re-check next audit). Scaleway adds GLM-5.3, DeepSeek V4.1 Flash and brings back Qwen3 Coder 30B. Pollinations adds GPT-5.4 (1M), Grok 4.6, Grok 4.3, Qwen3.8 2.4T A95B, Mistral Large 3 and Nemotron 3 Ultra. LLM7 gets MiniMax M2.7 back. Vercel AI Gateway adds Nemotron 3.5 Lightning Free and the stealth cat-jam slot. onomeo swaps space-bunny-alpha for the new stealth glyph-cluster.
+- ➖ **Dead or paid-flipped models removed**: NVIDIA killed Kimi K3 (deprecated 2026-10-14, gone from /v1/models). Novita's free tier collapsed from 8 to 2 models (the Ling free promo and GLM/Qwen/MiniMax free tiers all ended; Apodex 1.1 Mini moved in from OpenRouter as its replacement). Pollinations moved 5 models behind paid Pollen (qwen-coder, glm-5.2, grok-4.7, step-5-preview, nova-2-lite). OpenCode Zen rotated out 4 dead promo slots (step-5-preview-free, ling-3.1/3.0-flash-free, muse-spark-1.3-contributor). Kilo lost step-5-preview-free and ling-3.1-flash (both now paid). Vercel delisted ling-3.1-flash-free. onomeo delisted 5 models (gemini-3.8-flash, glm-5.3-flash, glm-4.7-flash, glm-4.6v-flash, space-bunny-alpha). Mistral retired Leanstral 1.5.1.
+- 🗑️ **Z.ai provider retired from the catalog**: the GLM Coding Plan free window ended 2026-10-07 and the plan is now paid-only (Lite $18/mo+), so the provider has zero free models left. The `zai` export stays as an empty array for backwards compat (same pattern as GitHub Models) and ZAI_API_KEY support remains in the code for existing configs. Free GLM-5.3-Flash lives on via NVIDIA NIM, OpenRouter, Pollinations, OrcaRouter and onomeo. 24 active providers now.
+- 🔧 **Context windows and ids fixed against live APIs**: 18 corrections, including the Requesty silent rename `novita/inclusionai/ling-3.1-flash` → `novita/ling-3.1-flash`, Ollama glm-5.2 976k → 1M and nemotron-3-nano 1M → 256k, LLM7 kimi-k2.6 262k → 256k, Mistral ministral-3b 131k → 128k, plus 7 Requesty and 3 Pollinations ctx normalizations.
+- ⏳ **OVHcloud Mistral Small 3.2 gets an auto-expiry date** (2026-11-04): OVH announced its decommission on 2026-08-04; the model still answers today so it stays listed, and the new `deprecatedAfter` tuple element removes it from the catalog automatically after the shutdown date.
+- 📊 **Quota labels refreshed**: Cerebras is now honest about its free tier ("$5 credits / 30 days, card required" instead of "generous dev tier"), onomeo's rate limit dropped to 30 req/5h, Novita shows "2 zero-price models left".
+- 🧪 **Tests updated for the catalog rotation** (same pattern as previous audits): router/installer fixtures moved from the dead `moonshotai/kimi-k3` NVIDIA slot to `z-ai/glm-5.3`, and the daemon-level zai normalizer subtest was folded into the existing unit coverage since the provider left the free catalog. 1185 tests pass.
+
+### Fixed
+
+- Nothing else; this release is a pure catalog audit.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "free-coding-models",
-  "version": "0.5.97",
+  "version": "0.5.99",
   "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",