vava-nessa/free-coding-models v0.5.64

on GitHub

Added

• 🔍 Full provider audit — every one of the 20 active providers was re-verified against live API endpoints and official docs. Researchers fetched current catalogs and reported structured JSON diffs (removed / added / fixed / confirmed). The agent then applied all safe changes to sources.js and audit_state.json.

  • 8 new free-tier models added across 7 providers:
    • opencode-zen — poolside/laguna-s-2.1-free (Laguna S 2.1 Free), ling-3.0-flash-free (Ling-3.0-flash Free)
    • openrouter — inclusionai/ling-3.0-flash:free (Ling-3.0 Flash), poolside/laguna-s-2.1:free (Poolside Laguna S 2.1), openai/gpt-oss-20b:free (GPT OSS 20B)
    • googleai — gemini-3.6-flash (Gemini 3.6 Flash), gemini-3.5-flash-lite (Gemini 3.5 Flash Lite)
    • qwen — qwen3.7-flash (Qwen3.7 Flash)
    • scaleway — gemma-4-31b-it (Gemma 4 31B IT)
    • llm7 — gemini-3.1-flash-lite (Gemini 3.1 Flash Lite)
    • ollama-cloud — kimi-k3 (Kimi K3), gemini-3-flash-preview (Gemini 3 Flash Preview), mistral-large-3:675b-cloud (Mistral Large 3 675B Cloud), qwen3.5 (Qwen 3.5 Cloud)

• 🧪 New audit_state.json fingerprint system — every provider audit now records lastAudited + an MD5 fingerprint of its confirmed + added model IDs. Next audit skips a provider automatically when its fingerprint hasn't changed.

• 📜 Migrated provider_updates/ archival to git history — the v1 skill kept per-provider markdown reports + per-version archives. v2 keeps git as the single source of truth: every sources.js change ships in a commit with a descriptive message, so git log -- sources.js shows the full audit trail.

Changed

• 📊 Catalog size 231 → 222 models (net -9). Free-tier shrinkage on OpenRouter (23 → 15) was partly offset by additions on opencode-zen, googleai, qwen, ollama-cloud, llm7, scaleway.
• 🔧 Provider-level fixes:
  • nvidia — nvidia/mistral-nemotron → mistralai/mistral-nemotron (namespace change)
  • cloudflare — @cf/deepseek/... → @cf/deepseek-ai/..., @cf/ibm/... → @cf/ibm-granite/... (namespace changes); Gemma 4 26B ctx 128k → 256k
  • ollama-cloud — 9 ctx corrections: glm-5.2, minimax-m3, deepseek-v4-flash, deepseek-v4-pro upgraded to 1M; gemma4:31b downgraded to 128k; kimi-k2.6, kimi-k2.7-code to 256k; glm-5.1, minimax-m2.5 to 198k
  • scaleway — 4 ctx corrections (Dedicated-tier vs Serverless-tier): Mistral Medium 3.5 256k → 180k (Serverless), Llama 3.3 70B 100k → 128k (Dedicated), Devstral 2 200k → 260k (Dedicated), GLM 5.2 256k → 1M
  • sambanova — standardize 5 entries to 131k (exact API value 131072); MiniMax-M2.7 192k → 196k
  • cerebras — 3 ctx corrections to actual free-tier limits: zai-glm-4.7 128k → 64k, gpt-oss-120b 128k → 65k, gemma-4-31b 128k → 65k
  • zai — zai/glm-4.7-flash 203k → 200k
  • codestral — codestral-2508 256k → 128k
  • novita — tencent/hy3 262k → 256k
  • ovhcloud — Qwen3-Embedding-8B ctx '-' → '32k'
• 📈 README updated to reflect new model count — badges, prose, /stats sample shape, and benchmark copy now say ~222 models instead of ~191.

Removed

• 🗑️ 31 dead models removed from the live catalog (kept in git history for posterity):
  • nvidia (14): MiniMax M2.7, Llama 4 Maverick, Mistral Small 4, Ministral 14B, Seed OSS 36B, Stockmark 100B, Qwen2.5 Coder 32B, Phi 4 Mini, Mistral Large 675B, DeepSeek V3.2, DeepSeek R1, Mistral Nemotron (old ns), Nemotron 3 Nano (old id), Llama 4 Scout
  • openrouter (11): Poolside Laguna XS.2, Tencent Hy3, Qwen3 Coder, Llama 3.3 70B Instruct, Hermes 3 Llama 3.1 405B, Qwen3 Next 80B A3B, OpenRouter Free, Llama 3.2 3B Instruct, Dolphin Mistral 24B Venice, LFM 2.5 1.2B Instruct, LFM 2.5 1.2B Thinking
  • ovhcloud (3): Qwen3 Coder 30B MoE, Qwen3 32B, Mistral 7B Instruct v0.3
  • opencode-zen (1): Tencent Hy3 Free
  • novita (1): Qwen3.5 Plus
  • routeway (1): Poolside Laguna M.1 (model available: false)

Skipped (flagged for review)

• ⏸️ llm7 subagent reported minimax-m2.6 removed — but our catalog has minimax-m2.7, not m2.6. The subagent's "removal" doesn't match any entry in sources.js, so it was discarded. m2.7 stays in the catalog until directly verified.
• ⏸️ routeway 2 ID "fixes" flagged as paid — deepseek-v4-flash:free and step-3.5-flash:free are listed as free in our catalog, but the live API shows they have non-zero prices. Skipped for this release; vava will decide whether to keep them in the free catalog or move them out.

Maintenance

• 🧪 779 tests pass (pnpm test). One test updated (test/test.js:989) to reflect the new NVIDIA catalog head — minimaxai/minimax-m2.7 is gone, so the test now asserts the next 5 models in the discovery-ordered list.
• 🛡️ pnpm start runs without runtime error. TUI / Web Dashboard / daemon all load cleanly with the new 222-model catalog.
• 📦 audit_state.json refreshed — every provider's lastAudited is now 2026-07-27 and each has a fresh fingerprint for the next audit cycle.

Files

• Modified: sources.js (30 net lines: -31 removed, +21 added, ~30 fixed inline), audit_state.json (timestamps + fingerprints refreshed), README.md (6 references to 191 → 222), test/test.js (1 test updated), package.json (version bump 0.5.63 → 0.5.64).
• New: changelog/v0.5.64.md, audit_state.json (refreshed), .pi/skills/update_models/SKILL.md (v2 skill), 20 /tmp/audit_diffs/<provider>.json (transient, gitignored).

Breaking changes

None. All fixes are in-place context-window or namespace corrections. Removed models are still cited in git log -- sources.js and the v0.5.64 changelog above for migration reference.