vava-nessa/free-coding-models v0.6.1

on GitHub

Full provider audit: all 25 tracked catalogs checked against their live APIs and official docs, one researcher per provider. 38 model rows removed, 39 added, 8 corrected. Net catalog: 258 models across 24 providers (onomeo exits the catalog entirely).

Removed

• onomeo (whole provider, 15 models): the provider removed its ENTIRE free-model lineup (8 tracked models appear in its own "This week's changes" list, the rest were already delisted). The current onomeo catalog is 100% paid, and the only no-charge access left is browser-chat rewards that do not work via API. The provider is retired from the catalog following the github-models/zai pattern: ONOMEO_API_KEY config support stays for backwards compatibility, but no onomeo rows are exposed anymore.
• OpenRouter: poolside/laguna-xs-2.1:free and poolside/laguna-s-2.1:free are fully delisted (zero poolside entries left in the live catalog, free or paid) and nvidia/nemotron-3-super-120b-a12b:free lost its free variant (the paid model remains). Replacement on the free shelf: recursive/flint-v1.0:free.
• Kilo: the same Poolside Laguna pair plus nvidia/nemotron-3-super-120b-a12b:free were removed from the gateway, mirroring OpenRouter's delisting.
• OpenCode Zen: 4 free models died since the last audit: nemotron-3-ultra-free, space-bunny-free, longcat-2.5-preview-free, exo-free (absent from the live /v1/models, completion probes return the same upstream-unavailable error as a bogus id; the docs table is stale on nemotron-3-ultra-free). jev-1.13-free stays live even though it is hidden from the public listing.
• Google AI: gemini-3.7-flash and gemini-3.5-flash were deprecated on 2026-10-08 and now auto-route to 3.8/3.6 Flash, so they were duplicate entries and are removed. The gemini-2.5 trio (2.5-flash, 2.5-flash-lite, 2.5-pro) was deprecated on 2026-11-02 and closes for new users, but still serves existing keys until 2026-12-01: they are kept with a deprecatedAfter: 2026-12-01 sunset so the catalog auto-filters them on Dec 1 instead of breaking them today.
• Scaleway: glm-5.2 is deprecated (EOL 2026-11-16, requests already route to glm-5.3) and qwen3-coder-30b-a3b-instruct lost its Serverless access again (EOL 2026-10-01 re-confirmed, reversing the previous re-add).
• LLM7: kimi-k2.6 and kimi-k2.7-code flipped to paid-only (tier=pro, usage_based_only=true; a free token now gets 401). Kimi K2.6 remains free on NVIDIA NIM, Pollinations and Ollama Cloud.
• Routeway: the deepseek-v4-flash:free and minimax-m2.7:free slots rotated to the newer generations (see Added).
• Mistral LP: mistral-medium-3-5 was officially deprecated on 2026-10-31 (retirement 2027-04-30, designated replacement Mistral Large 4), and mistral-vibe-cli-latest now resolves to that same deprecated Medium line, so both are removed until Mistral repoints the alias.
• Vercel AI Gateway: poolside/laguna-s-2.1-free is delisted (only the paid laguna-s-2.1 remains) and the stealth/cat-jam stealth slot rotated out.
• NVIDIA NIM: z-ai/glm-5.3 reached end of life on 2026-10-28 (API returns HTTP 410 Gone); z-ai/glm-5.3-flash remains.

Added (39)

• NVIDIA NIM (4): moonshotai/kimi-k2.6 (SWE-bench Verified 80.2%, the new top score in the catalog, ahead of Nemotron 3 Ultra's 71.9%), minimaxai/minimax-m3.1-flash (DeepSWE 73.7%), mistralai/mistral-large-4-1t-a52b and nvidia/nemotron-3.5-bolt-e4b.
• Pollinations (5): openai/gpt-6.1-terra, openai/gpt-5.6-sol, openai/gpt-5.6-terra, stealth/glyph-cluster (experimental coding stealth slot) and meta/llama-3.3-70b-instruct.
• Ollama Cloud (4): minimax-m3.1-flash (976k ctx), muse-spark-open, mimo-v2.6-pro, mimo-v2.6-flash.
• Alibaba DashScope (3): third-party models now hosted on the international Model Studio: glm-5.3, kimi-k2.7-code, deepseek-v4.1-flash.
• SiliconFlow (6): the free tier re-expanded. Qwen/Qwen3-8B, deepseek-ai/DeepSeek-R1-0528-Qwen3-8B, THUDM/GLM-4-9B-0414, THUDM/GLM-Z1-9B-0414, Qwen/Qwen3.5-4B and Qwen/Qwen2.5-7B-Instruct are all back at ¥0 on the official pricing page, reversing the 2026-09 paid/deprecated verdicts. zai-org/GLM-5.3-Flash stays excluded (listed paid on the rendered pricing table).
• Novita (3): bunny (the renamed Space Bunny stealth model, now verified zero-price), plus two new 1M-ctx gateway routers launched 2026-10-27: nvrouter/auto and nvrouter/fusion.
• Kilo (3): thinkingmachines/inkling:free (the flagship is back as a $0 entry), recursive/flint-v1.0:free and google/gemma-4-31b-it:free.
• OpenCode Zen (2): flint-40b-free (Recursive Flint, 40B MoE coding model, 74.2% SWE-bench Verified) and pulse-alpha-free (new stealth, 1M ctx, geo-restricted outside the US and allied regions).
• OpenRouter (1): recursive/flint-v1.0:free (S+, 74.2%, 262k ctx), the new free coding pick filling the Laguna slots.
• Routeway (2): deepseek-v4.1-flash:free and minimax-m3.1-flash:free, both $0/$0 at 42k ctx.
• Mistral LP (1): mistral-large-4 (1T MoE, 1M ctx), verified answering on a free Experiment key; tier S by DeepSWE (61.7%) convention.
• SambaNova (1): MiniMax-M3.1-Flash (Preview, 1M ctx), filling the slot MiniMax M2.7 left.
• OrcaRouter (1): deepseek/deepseek-v4.1-flash-free (1M ctx, 74.2 DeepSWE).
• Vercel AI Gateway (1): inclusionai/ling-3.1-flash-1019 ($0/$0, 1M ctx), reopening the Ling free slot. The base reflection/beam id also went $0 but is a duplicate of the tracked beam-free, so it was not added.
• Cloudflare (1): @cf/meta/llama-4-scout-17b-16e-instruct re-added (tier C). Every new 2026 flagship on Cloudflare is Workers-Paid-only since 2026-07-28, so nothing stronger was eligible.
• Runcrate (1): deepseek-ai/DeepSeek-V4.1-Flash-1017, a dated V4.1 Flash snapshot now listed as its own free-tier id (first audit since the provider landed).

Fixed

• Google AI SWE scores filled from the Vals.ai leaderboard: gemini-3.8-flash 80.0%, gemini-3.6-flash 79.6%, gemini-3.5-flash-lite 75.0%, gemini-3-flash-preview 78.0% → 75.0%. gemini-3.1-flash-lite's 2027-05-07 shutdown is now wired into deprecatedAfter so the auto-filter handles it.
• ollama-cloud: minimax-m3 context corrected 512k → 500k (live /v1/models reports 512000).
• requesty: poolside/laguna-xs.2 score corrected 70.9% → 69.9% (70.9% belongs to successor Laguna XS 2.1; 69.9% is the official Poolside model card for XS.2).
• ovhcloud: bge-multilingual-gemma2 end-of-service 2027-01-21 announced by OVHcloud, wired into deprecatedAfter.
• codestral: the model card now shows 128k ctx while OpenRouter and prior live reads still show 256k; the contested change was NOT applied, the catalog keeps 256k until a keyed re-check settles it.

Docs & tooling

• README provider table and docs/providers.md regenerated for the new catalog shape (24 providers / 258 models).
• audit_state.json refreshed: all 24 providers re-fingerprinted, onomeo's entry dropped, audit date 2026-11-03.
• Test fixtures that pinned removed ids migrated to their live successors: router/endpoint-installer fixtures moved from z-ai/glm-5.3 to z-ai/glm-5.3-flash, and the Mistral temperature-clamp router test moved from mistral-medium-3-5 to mistral-large-4 (the router only routes to models present in the catalog, so fixtures must track live ids).