decisionmodels releases s1-llm-auto-router, a 183 MB CPU router under MIT
A pruned mmBERT-small that answers seven routing questions in one CPU pass. The card reports 0.898 category accuracy on its held-out set.
The Hugging Face account decisionmodels published s1-llm-auto-router on Hugging Face on 15 October 2026, by repository creation date. The card describes a decision model fine-tuned to classify a request to an LLM before you route it. It answers seven questions in one CPU pass: category, difficulty, stakes, and four yes/no flags for tools, vision, long context and follow-up.
It is built on jhu-clsp/mmBERT-small, pruned to 11 layers and distilled from a fine-tuned mmBERT-base teacher. The ONNX file is 183 MB with int8 embeddings and seven heads, and runs without PyTorch. The licence is MIT. A hosted API is listed at USD 0.001 per million input tokens, with a note that it may change.
On a held-out test of 2,175 examples the card reports category accuracy 0.898, all seven answers exact 0.819 and routing agreement 0.742, with about 26 ms p50 on four CPU threads. On an independent set of 100 requests labelled by Claude Opus, category is 0.88 and routing agreement 0.60. The card says that sample carries roughly plus or minus 8 points and that the maker built the model.
Training data is about 39k synthetic requests, with no real user traffic. The card says so.
Useful if you route a lot and want it off the GPU.
Sources
Editor's first pass (AI-assisted)
- Real 4: MIT weights as an ONNX file, plus a hosted API with listed pricing.
- Useful 3.5: Classifies a request by category, difficulty, stakes and four flags before choosing a model.
- Notable 3: Aimed squarely at LLM routing, and runs on CPU.
- Novel 3: Distilled, int8, no PyTorch needed.
Comments