HuatuoGPT-o1-70B

HuatuoGPT-o1-70B is a medical reasoning model from the FreedomIntelligence team (CUHK-Shenzhen), released in December 2024. It takes LLaMA-3.1-70B and trains it for complex medical reasoning in two stages: supervised fine-tuning on medical chain-of-thought data, then reinforcement learning with verifier-based rewards to sharpen the reasoning further. The model answers in an explicit thinks-before-it-answers format, with a ## Thinking section followed by a ## Final Response, in the style of the o1 models.

The team open-sourced the models, training data and code, which is more than most medical-model releases offer, and shipped four sizes: 8B and 70B on LLaMA-3.1 (English) plus 7B and 72B on Qwen2.5 (English and Chinese). The paper reports gains over the LLaMA-3.1-70B backbone on medical reasoning benchmarks. Adoption has been thin (around a hundred downloads) and there has been no follow-up release, so treat it as a research artifact: interesting method, honest open release, but no community validation to lean on. As with all medical models, the evals are English-centric and nothing here is clinically validated.

• Authors report HuatuoGPT-o1 outperforms its LLaMA-3.1-70B backbone on complex medical reasoning benchmarks after SFT plus verifier-reward RL (provider claim, paper)
• No independent benchmark table verified; medical evals are English-centric