v0.5.50

Changed

• Full model catalog audit — verified all 20 providers against their official docs and APIs (2026-07-18)
• Auto-deprecation support — models with a deprecatedAfter date in the tuple are now automatically filtered from MODELS after that date, without requiring a code change
• Cerebras — zai-glm-4.7 marked as ⚠️ DEPRECATED shutdown 2026-08-17 per official Cerebras notice; will disappear automatically after that date

Fixed

• NVIDIA NIM — removed magistral-small-2506 (not found on NIM) and qwen3-coder-480b (deprecated free endpoint); fixed context windows: mistral-medium-3.5-128b 128k→256k, nemotron-3-nano-omni 128k→256k, mistral-nemotron -→128k; added deepseek-r1 at A tier
• Codestral — fixed codestral-2508 context window 128k→256k
• OpenRouter — removed gpt-oss-120b:free and gpt-oss-20b:free (no longer available on free tier)
• Ollama Cloud — full catalog refresh: 19 outdated models removed, 10 new models added (kimi-k2.6, kimi-k2.7-code, kimi-k2.5, deepseek-v4-flash, deepseek-v4-pro, glm-5.1, glm-5.2, minimax-m2.7, and more); catalog now verified against official ollama.com/search?c=cloud
• Novita AI — removed stale models (qwen3.6-plus, nex-agi/nex-n2-pro, minimax/m2-her); added tencent/hy3
• LLM7 — removed devstral-small-2:24b; added minimax-m2.7 (S+) and gpt-oss:20b (A+)
• OpenCode Zen — added hy3-free (Tencent Hy3 Free)

Confirmed (no changes)

• groq, googleai, github-models, mistral, cloudflare, sambanova, ovhcloud, zai, scaleway, qwen, kilo, routeway — all models verified current

v0.5.49

Full Changelog: v0.5.47...v0.5.49

v0.5.47

Added

• New BETA fcm-opencode plugin — an OpenCode adapter that brings free-coding-models scans, routing, and provider injection into OpenCode. It reuses the exact same scan / ranking / API-key / daemon / model-safety code as fcm-pi, so provider fixes and safety filters stay in one place. Install via a local symlink (see opencode-plugin/README.md). Commands:
  • /fcm — scan and list ranked choices without switching
  • /fcm 1 / /fcm 2 / … — explicitly switch OpenCode config to ranked model #N
  • /fcm best — explicitly switch to the best ranked model
  • /fcm rescan — force a fresh scan
  • /fcm status (or /fcm-status) — plugin diagnostics (cache age, daemon, active model)
  • /fcm router (or /fcm-router) — explicitly switch OpenCode to the local FCM Smart Router daemon
• New shared pi-extension/lib/pi-model-config.js module — centralizes the Pi model safety policy so both adapters apply identical rules: parseContextWindow (8k / 128k / 1M → number), getPiMaxTokens (caps completions to leave prompt room), isPiContextUsable (hides tiny-context models that pass a probe but fail real agent prompts, e.g. 8k Cerebras), and getPiReasoningFlag (keeps OpenAI-compatible FCM models in non-thinking mode).
• runFcmScan({ mode }) now supports auto, daemon, and direct scan strategies. OpenCode uses daemon at startup so direct probes never block boot.

Changed

• FCM-Pi is now silent by default. No startup scan, no footer noise, and no automatic model switch on Pi boot or /resume. Run /fcm to scan and pick a model explicitly.
• Error-triggered picker replaces silent auto-failover. When the active model returns HTTP 4xx/5xx, FCM reopens the picker and marks the failing model 🔴 BUGGED instead of switching automatically — you stay in control of which model runs.
• Pi scan-progress footer now shows the branded > free-coding-models badge in the exact same green/white-on-black colours as the main TUI header logo, instead of scrolling live model names. The live % progress and (completed/total) counter stay beside it.
• OpenCode startup is light by default: fresh cache first, daemon second, no direct scan unless the user explicitly runs /fcm. Long probes never freeze OpenCode boot.
• README updated with a new fcm-opencode section and a refreshed FCM-Pi section reflecting the silent-by-default behaviour.

Notes

• fcm-pi and fcm-opencode remain BETA and are installed via local path / symlink, not via the npm free-coding-models package.
• The full shared-core extraction into packages/fcm-agent-core / packages/fcm-opencode is still in progress (tracked in .kandown/tasks/t2.md). This release ships the working adapters and the extracted model-safety module ahead of that restructure.

v0.5.46

Fixed

• LLM7 catalog corrected against live /v1/models API. The previous audit had added several models that are NOT free on LLM7. The live API shows that of LLM7's 12 models, only 2 are on the turbo (free, noKeyNeeded) tier; everything else — including the claude-* (Fable 5, Opus 4.8, Sonnet 5), gpt-5.4 / gpt-5.4-mini / gpt-5.5, glm-5.2, kimi-k2.6, deepseek-v4-flash, and minimax-m2.7 entries — are pro (usage-based paid) and were incorrectly listed as free.
• Removed 10 non-free (pro) models from LLM7, keeping only the two genuinely-free turbo models: devstral-small-2:24b (S / 68.0%) and codestral-latest (A / 40.0%). Total catalog: 257 → 247 models.
• Removed the 1.05M context-window entries (gpt-5.5 / gpt-5.4) that were causing alignment drift; no 1.05M remains in the catalog.

Notes

• LLM7's free (turbo) tier now has only 2 coding models. The llm7 provider remains noKeyNeeded: true and should only surface free models per the source-header rule ("Pro-tagged models from /v1/models are intentionally excluded").

v0.5.45

Changed

• SWE-bench Verified scores refreshed for all 257 models. 12 web-research agents audited every model family against the SWE-bench Verified leaderboard (swebench.com), SWE-Rebench, Artificial Analysis, and official model cards.
• 68 models went from - (unknown) to a real SWE-bench Verified score (e.g. GPT-4.1 → 54.6%, GPT-5 → 74.9%, DeepSeek V4 Pro → 80.6%, Qwen3.7 Max → 80.4%, Gemini 2.5 Pro → 63.8%, GLM-5.2 → 82.8%, Kimi K2.6 → 80.2%, Llama 4 Maverick → 74.8%, Nemotron 3 Ultra → 71.9%).
• Tiers recomputed from real SWE-bench Verified scores using the project scale (S+ ≥70%, S 60–70%, A+ 50–60%, A 40–50%, A- 35–40%, B+ 30–35%, B 20–30%, C <20%). Several models dropped tiers to match verified numbers (e.g. Llama 3.3 70B A- → B at 22.0%; GPT-OSS-20B A → B+ at 50.3%; Mistral Small 4 S → A- after its public score was not found, kept manual estimate).
• Existing scores corrected where research disagreed with prior self-reported/manual values (e.g. GPT-OSS-120B 60% → 62.4%; Qwen3.6 Plus 72% → 78.8%; MiniMax M2.7 80.2% confirmed; Gemini 3.1 Pro 78% → 80.6%).

Notes

• 55 models remain - (no public SWE-bench Verified score): embeddings (bge, Qwen3-Embedding), vision-only (Pixtral, Llama vision, GLM-4.6V), small/edge models (Phi-4, Ministral 3B, Granite, Gemma 3 4B/12B), routers (Kilo, OpenRouter Free, Mimo, Nex, RNJ, Big Pickle, Poolside where not listed), reasoning-only (Nemotron Nano Omni), and unverifiable/fictionally-named entries (Claude Fable 5, Claude Sonnet 5 returned scores but are unverified; Claude Opus 4.8 returned 88.6% but is unverified). Their tiers retain the prior manual estimates.
• Source scores are self-reported / leaderboard-aggregated and drift over time; re-run the audit periodically.

v0.5.44

Added

• Provider catalog audit (all 20 providers) — refreshed every provider against live official docs/endpoints via the update_models skill.
• NVIDIA NIM: added Qwen3 Coder 480B, Mistral Nemotron, Llama 3.3 Nemotron Super 49B v1.5, DeepSeek V3.2, Qwen2.5 Coder 32B, Magistral Small 2506, Nemotron 3 Nano, Llama 4 Scout (A+ / 10M ctx); GLM-5.1 → GLM-5.2.
• GitHub Models: added GPT-5 / GPT-5 Mini / GPT-5 Nano / GPT-5 Chat, GPT-4o / GPT-4o Mini, o3 / o4-mini / o3-mini / o1 / o1-mini / o1-preview, DeepSeek-R1 / R1-0528, Cohere Command A, Phi-4 family (20 models).
• Cerebras: added Gemma 4 31B.
• Mistral: added Ministral 3 family (14B / 8B / 3B); Magistral Medium 1.2 retired.
• Cloudflare: added QwQ-32B, DeepSeek-R1-Distill-Qwen-32B, Llama 3.1 8B Instruct (Fast); removed deprecated Kimi K2.5 and Mistral 7B v0.2.
• OpenRouter: added Tencent Hy3, Qwen3 Coder, Poolside Laguna XS 2.1, Llama 3.3 70B, Hermes 3 405B, Qwen3 Next 80B, Llama 3.2 3B, Dolphin Mistral 24B, Nemotron 3.5 Content Safety, LFM 2.5 1.2B (instruct + thinking); removed Owl Alpha; renamed Nemotron Nano 30B → Nemotron 3 Nano 30B.
• Codestral: added Codestral 2.
• Z.ai: added GLM-4.6V-Flash; GLM-4.7-Flash ctx 200k → 203k.
• Scaleway: added GLM 5.2, Mistral Medium 3.5 128B, Gemma 4 26B MoE, Pixtral 12B; dropped mistral/ prefix on Mistral Large 675B id.
• Qwen: added Qwen3.7 Plus, Qwen3.6 Max Preview, Qwen3 Coder 480B, Qwen3.5 397B/122B/35B/27B, Qwen3 Next 80B (thinking + instruct), Qwen3 30B; Qwen3 235B & Qwen3 32B ctx corrected 256k → 128k.
• LLM7: added Claude Opus 4.8 / Sonnet 5 / Fable 5, GPT-5.5 / GPT-5.4 / GPT-5.4-mini, GLM 5.2, MiniMax M2.7; removed Qwen3 235B & Mistral Small 3.2; ctx fixes on DeepSeek V4 Flash (1M), Kimi K2.6 (240k), Devstral Small 2 (255k).
• Ollama Cloud: removed 7 paid/locked models (MiniMax M2.7, GLM-5.1, Kimi K2.6, DeepSeek V4 Pro/Flash, Qwen3.5 397B, Mistral Large 3); added 17 free-tier models (Gemma 3 4B/12B/27B, MiniMax M2/M2.1/M2.5, Ministral 3 3B/8B/14B, GLM 4.6, Cogito 2.1 671B, Devstral Small 2 24B, Nemotron 3 Nano 30B, Qwen3 Next 80B, Qwen3 VL 235B + instruct, RNJ 1 8B); ctx fixes on Devstral 2, Nemotron 3 Ultra/Super, GLM-4.7, MiniMax M3.
• Novita: corrected Qwen3.6 Plus SWE score (78.8%) and MiniMax M2 HER ctx (66k).
• Routeway: removed MiniMax M2 and Mistral Nemo free tiers.

Fixed

• Cerebras: corrected both model context windows from 8k → 128k.
• GitHub Models: Llama 4 Scout ctx 1M → 10M.
• NVIDIA NIM: Kimi K2.6 ctx 256k → 262k; Nemotron 3 Nano 30B ctx 128k → 1M.

Changed

• audit_state.json regenerated for all 20 providers (total catalog: 257 models, up from 157).
• Catalog arrays now grouped by tier with inline tier comments for easier review.

Notes

• Low/medium-confidence sources (Ollama Cloud third-party probe, Routeway chimera/umbra, LLM7 guessed tiers, Novita free-status) were applied as-is per request; verify before relying on them in production.

v0.5.43

Added

• Pi Extension documentation: Added a full ## π Pi Extension — FCM-Pi section to the main README.md covering installation, scan progress display, commands, composite ranking, and context window notes.
• Rewrote pi-extension/README.md: Complete documentation rewrite covering all current features — animated progress scan, provider-in-status-bar, auto-failover, daemon integration, file structure, and provider-specific notes (Cerebras limits, NVIDIA NIM quota).

v0.5.42

Added

• Added an animated magenta progress spinner in the Pi TUI status line during candidate scans.
• Added live rolling displays of the specific model and provider currently being probed or benchmarked.
• Integrated the provider name (e.g., Cerebras, Nvidia) directly into the active model name shown in the Pi status line.

v0.5.41

Changed

• Adjusted the context window size of Cerebras models (zai-glm-4.7 and gpt-oss-120b) to 8k in the model catalog (sources.js) to reflect the strict 8192 total token limit of the Cerebras free tier API.

Fixed

• Added a custom compatibility flag "supportsTools": false in openai-completions.js to allow completely disabling tool parameter injection for models with very small context windows.

v0.5.40

Fixed

• Avoid sending thinking: { type: 'disabled' } to Cerebras, Mistral, Groq, and SambaNova during ping probes, preventing immediate 400 errors and redundant retry calls.