feat: sync models from router (+10 added, 6 reasoning-capable, 6 artifacts, -0 removed)#2403
Conversation
…facts, -0 removed) Add descriptions for new router models missing from prod.yaml/dev.yaml: - XiaomiMiMo/MiMo-V2.5-Pro (1T MoE, reasoning, artifacts) - moonshotai/Kimi-K2-Thinking (1T MoE, reasoning, artifacts) - nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 (550B MoE, reasoning, artifacts) - stepfun-ai/Step-3.7-Flash (198B vision-language MoE, reasoning, artifacts) - deepreinforce-ai/Ornith-1.0-35B (35B MoE, reasoning, artifacts) - deepreinforce-ai/Ornith-1.0-35B-FP8 (35B MoE, reasoning, artifacts) - google/gemma-3-12b-it, google/gemma-3-4b-it (dense multimodal, <32B) - microsoft/phi-4 (14B dense, <32B) - speakleash/Bielik-11B-v3.0-Instruct (11B Polish, <32B) Reasoning flags verified against live provider docs (DeepInfra/Novita reasoning_effort; StepFun low/medium/high). No deprecated models to prune. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PHnvbW9eLvNuMPsCkwGR3o
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5fab139b88
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| { "id": "deepreinforce-ai/Ornith-1.0-35B", "description": "RL-tuned 35B MoE coding agent with thinking mode and 262K context.", "supportsReasoning": true, "supportsArtifacts": true }, | ||
| { "id": "deepreinforce-ai/Ornith-1.0-35B-FP8", "description": "FP8 Ornith 35B MoE for efficient agentic coding with thinking mode.", "supportsReasoning": true, "supportsArtifacts": true }, |
There was a problem hiding this comment.
These two overrides mark Ornith as supportsReasoning, which in this app specifically means the model accepts a low/medium/high reasoning_effort knob (src/lib/server/models.ts), so the chat UI exposes the Effort selector and forwards the selected value. Ornith's HF/DeepInfra docs describe it as always emitting a <think>/reasoning_content trace via a parser, not as supporting configurable effort levels; when users choose Low/Medium/High, HuggingChat will send an unsupported reasoning_effort param for these models. Drop the flag or add provider-specific handling before enabling it; the same issue is present in chart/env/dev.yaml.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Confirmed — dropped supportsReasoning from both Ornith entries in prod.yaml and dev.yaml (commit 35866d0), keeping supportsArtifacts.
Verified against the HF model card: Ornith always opens with a <think>/reasoning_content trace via a parser but documents no reasoning_effort / configurable effort levels, so it doesn't match this app's supportsReasoning semantics (accepts a low/medium/high reasoning_effort knob). One small correction to the impact: DeepInfra's reasoning docs say reasoning_effort on a model that doesn't support it "has no effect" (silently ignored, not an error), so the practical bug was an inert Effort selector rather than a broken request — but removing the flag is still the right call.
For reference, the other new reasoning flags were re-checked and do document reasoning_effort: Step-3.7-Flash (StepFun docs), MiMo-V2.5-Pro (AI/ML API docs), and Kimi-K2-Thinking (effort levels), so those stay.
Generated by Claude Code
Ornith always emits a <think>/reasoning_content trace via a parser but its HF model card documents no reasoning_effort / configurable effort levels. supportsReasoning in this app specifically means the model accepts a low/medium/high reasoning_effort knob (src/lib/server/models.ts), which the UI would then expose and forward. Removing the flag so the inert Effort selector no longer appears for these two entries. supportsArtifacts kept (35B >= 32B). Applies to both prod.yaml and dev.yaml. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PHnvbW9eLvNuMPsCkwGR3o
2 participants
Add descriptions for new router models missing from prod.yaml/dev.yaml:
Reasoning flags verified against live provider docs (DeepInfra/Novita
reasoning_effort; StepFun low/medium/high). No deprecated models to prune.
Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01PHnvbW9eLvNuMPsCkwGR3o