Porting high-quality chain-of-thought behavior into an open, runnable model matters because most frontier reasoning models are locked behind proprietary APIs. This project distills Claude Opus 4.7’s explicit reasoning style into a permissively-licensed Mixture-of-Experts checkpoint so individual researchers and practitioners can run Claude-like stepwise reasoning locally or in self-hosted endpoints.
Key Capabilities
- Reasoning-style distillation: trained on ~7.8k high-quality Opus 4.7 reasoning traces to teach the model to emit explicit
<think>…</think>chains before final answers, improving traceability for multi-step problems. This is a style/behavior distillation (how to reason), not a factual knowledge injection. - MoE efficiency: base is a 35B-parameter MoE (256 experts, 8 routed + 1 shared) with only ~3B parameters active per token, enabling high-capacity reasoning at the inference cost closer to a small dense model. Full bf16 inference fits on a single 80GB A100/H100.
- Long-context and local/runtime options: native 64k context, recommended serving with vLLM for routing/KV cache benefits, and GGUF-quantized builds (IQ4_XS, Q5_K_M, Q8_0) for LM Studio / llama.cpp to run locally.
- Evaluated on standard benchmarks: strong CoT performance (GSM8K CoT reported 84.3% flexible-extract) and competitive MMLU-Pro (74.9%) on STEM-heavy tasks, while known extraction-format issues affected some AIME/GPQA runs (pending fixes).
Who it's for — tradeoffs and fit
Great fit if you need an open checkpoint that reproduces explicit chain-of-thought style for graduate-level STEM, contest math, long code reasoning, or agent planning and you can budget large generation lengths (tens of thousands of tokens). Look elsewhere if you need up-to-date world knowledge beyond the base model, low-latency short-turn chat, or a model guaranteed to match Anthropic’s factual priors — distillation transfers reasoning style, not new facts. Also note the fine-tune used attention-only LoRA (expert FFNs left intact), so some domain gaps versus the teacher may remain.
Where it sits in the stack
Practically, it’s a compromise: Claude-like reasoning cadence available under permissive licensing, built on a MoE base for capacity and on-ramps for both server (vLLM, HF endpoints) and local use (GGUF/llama.cpp). If you prioritize reproducible chain-of-thought outputs and the ability to self-host or audit long reasoning traces, this model is a rare, practical option.
