蚂蚁集团发布开源权重医疗 MoE 模型 LingSante-3.1-flash,总参数 560B、每 token 激活 25B,权重以 MIT 许可在 Hugging Face 开放。
原文交代了医疗数据与通用数据混合预训练、分技能 RL 再加多教师蒸馏的训练路径,可供关注垂直模型训练的人参考。
Ant Group released LingSante-3.1-flash, a 560B-parameter medical MoE model, and it tops GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon on HealthBench Professional, HLE, BrowseComp and EHR-Complex.
HealthBench Professional cover answering real clinician requests that real doctors bring to chatbots. EHR-Complex tests an agent querying a hospital records database.
weights on Huggingface, with MIT-license.
Ant mixed medical and general data from pretraining onward, trained medical reasoning, tool use, search and clinician-style behavior as separate skills with their own RL rewards on physician- and verifier-checked tasks, then merged those specialists into one model through multi-teacher on-policy distillation.
Today, we release LingSante-3.1-flash, an open-weight MoE model with 560B total parameters and 25B activations per token. It unifies diverse medical capabilities while retaining general capabilities, including complex medical reasoning, long-horizon workflow execution, and professional behavior alignment. LingSante-3.1-flash moves beyond medical QA toward real-world healthcare problem. The model leads on HealthBench Professional (65.72), HLE (text-only, with tools: 72.22) and BrowseComp (with context: 94.10), outperforming GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon on all three. LingSante-3.1-flash also performs well on professional long-horizon medical workflows, such as PhysicianBench (57) and EHR-Complex (74.40). Together, these results establish LingSante-3.1-flash as the leading open-weight models for healthcare. The weights are available. Explore LingSante and build with it. More below ↓在 X 查看被引用的帖子
来源:Rohan Paul · x.com