v0.5.99

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.