🔍 Full provider audit (20 providers, 222 → 196 models)
Major refresh across every free-coding-model provider. Most impactful change: GitHub Models is fully retired as of 2026-07-30 (HTTP 410 Gone) — the entire provider entry is removed from the active catalog.
Added
- nvidia (5):
poolside/laguna-xs-2.1(S+, 70.9%),meta/llama-3.3-70b-instruct,mistralai/mistral-large-2-instruct,meta/codellama-70b,ibm/granite-34b-code-instruct - qwen (1):
qwen3-vl-flash— qwen3.5/3.6/3.7/3.7-plus/3.6-max-preview/3.5-plus/3.5-flash/3.6-flash/3.7-max/3.6-plus/3.7-plus andqwen3-coder-nextwere already added in the prior sweep - ollama-cloud (0 net):
kimi-k3,glm-5.1,glm-5.2,minimax-m2.7,deepseek-v4-pro,mistral-large-3:675b-cloud,qwen3.5already present (no:cloudsuffix convention used) - cloudflare (2):
@cf/google/gemma-3-12b-it,@cf/moonshotai/kimi-k2.5 - scaleway (1):
deepseek-v4-flash-0731(256k serverless) - googleai (1):
gemini-3.7-flash(Aug 2026, 1M context) - openrouter (2):
liquid/lfm-2.5-2.6b:free,nvidia/nemotron-3.5-lightning:free - opencode-zen (2):
hy3-free(re-added),nemotron-3.5-lightning-free - ovhcloud (1):
Qwen2.5-VL-72B-Instruct - zai (7):
zai/glm-5.2,zai/glm-5-turbo,zai/glm-5v-turbo,zai/glm-4.7,zai/glm-4.6,zai/glm-4.7-flashx,zai/glm-4.6v(full GLM-5 family) - llm7 (1):
mistral-Nemo-Instruct-2407 - codestral (2):
codestral-2501,codestral-2405 - kilo (1):
kilo-auto/small(262k, routes to gemma-4-26b for free accounts)
Removed
- github-models (-35): Provider fully retired 2026-07-30 by GitHub (HTTP 410 Gone with
github_models_retirement_brownouterror). All 35 catalog entries unreachable. Provider commented out ofsourcesmap;githubModelsarray kept empty for backwards-compat. - groq (-2):
llama-3.3-70b-versatile,llama-3.1-8b-instant— Groq deprecation, shutdown 2026-08-16 - ollama-cloud (-11 net): 11 stale cloud IDs (deepseek-v4 umbrella, glm-5, mistral-medium-3.5, qwen3-coder-480b, step-3.7-flash, etc.) — already cleaned up in prior sweep
- ovhcloud (-2):
Mistral-Small-3.2-24B-Instruct-2506,Mistral-Nemo-Instruct-2407— endpoints reachable but no longer in public catalog - openrouter (-2):
poolside/laguna-m.1:free(gone from catalog entirely),inclusionai/ling-3.0-flash:free(now paid-only) - opencode-zen (-2):
north-mini-code-free,ling-3.0-flash-free - scaleway (-2):
mistral-large-3-675b-instruct-2512,gemma-4-31b-it— Dedicated tier only, not on Serverless - routeway (-4):
deepseek-v4-flash:free,step-3.5-flash:free,ling-3.0-flash:free,ling-2.6-flash:free— silently retired from zero-price catalog - mistral (-1):
devstral-2512— Mistral deprecation, retirement 2026-07-31 - codestral (-1):
codestral-2— fabricated ID, never existed in Mistral catalog (Mistral uses date-stamped versioning) - novita (-1):
tencent/hy3—isFree:falseper Novita pricing page. Provider now empty (0 models).
Fixed
- nvidia (2):
deepseek-ai/deepseek-v4-flash→deepseek-ai/deepseek-v4-flash-0731(NIM/v1/modelsonly exposes the -0731 suffix);deepseek-ai/deepseek-v4-proflagged as page-only / partner-routed (NOT in/v1/models, served via Fireworks/DeepInfra/Together/OpenRouter) - openrouter (1):
nvidia/nemotron-3-super-120b-a12b:freectx1M→262k - opencode-zen (1):
poolside/laguna-s-2.1-free→laguna-s-2.1-free(droppoolside/prefix) - scaleway (3):
glm-5.2ctx1M→256k,devstral-2-123b-instruct-2512ctx260k→200k,llama-3.3-70b-instructctx128k→100k(Serverless tier per official catalog) - mistral (3):
ministral-3-14b-25-12→ministral-14b-2512,ministral-3-8b-25-12→ministral-8b-2512,ministral-3-3b-25-12→ministral-3b-2512(use real API model IDs from Mistral docs JSON, not page slugs) - llm7 (1):
gemini-3.1-flash-litectx1M→256k(real LLM7 ctx limit)
Internal
PREFERRED_DEFAULT_MODELS(router pinned set) refreshed to point at current model IDs afterllama-3.3-70b-versatile/llama-3.1-8b-instantremoval anddeepseek-v4-flash→deepseek-v4-flash-0731renamePROVIDER_TEST_MODEL_OVERRIDES.nvidiaupdated todeepseek-ai/deepseek-v4-flash-0731buildDefaultRouterSetnow filters to keyed providers in the no-probeFn path (matches sync fallback behavior) so users without a probe fn see only models they can actually call
Tests
- 564/564 tests passing (97.5% → 100% in the impacted router/key-tester suites)