Skip to main content
​
Model catalog pruning · tool calling on qwen3.8-27b
12 models removed. The following ids are no longer served and now return 400 model_not_found:gpt-5.1 · gpt-5.6-sol · gpt-5.6-terra · mimo-v2.5-pro · glm-5.1 · kimi-k2.5 · glm-5 · glm-4.7 · minimax-m2.5 · qwen3-235b · nemotron-super-3-120b · gpt-oss-120bIf your client had one of these configured, pick a replacement from GET /v1/models — for reasoning workloads, deepseek-v4-pro, kimi-k2.6, and glm-5.2 are the closest fits.Tool calling enabled on qwen3.8-27b. The model now reports supports_tools: true and accepts the standard tools / tool_choice fields.
​
Status semantics · public context ceilings · docs launch
Model status criteria reworked. The fleet status badge (/api/stats/fleet, the models page) now judges health from error rates only:
  • degraded no longer triggers on slow first-byte times — reasoning models and long prompts no longer show as degraded while healthy.
  • down now means “a majority of recent requests failed”, never “no traffic”. Quiet models fall back to a 24-hour window before showing idle.
Input ceilings are now public. /api/stats/fleet model rows carry max_input_tokens — the same live, operator-tuned limit /v1/models reports — so unauthenticated surfaces can show real context ceilings.Documentation launched. OpenAPI reference for the six /v1 endpoints, the NovelAI-compatible Illustration API, guides, and this changelog.