ModelsSelected

蚂蚁集团发布 560B 参数医疗 MoE 模型 LingSante-3.1-flash,MIT 协议开源

Why we picked it

蚂蚁把医疗推理、工具调用、搜索拆成单独技能分别做 RL,再用蒸馏合并成 560B 的 MoE,权重还是 MIT 开源的。四项医疗基准分数压过了 GPT-6 Astra 和 Claude Opus 5.5,做医疗 Agent 或医院信息系统的开发者可以直接下载权重试试。

蚂蚁集团发布医疗专用 MoE 模型 LingSante-3.1-flash,参数量 560B,权重以 MIT 协议上传 Hugging Face。它在 HealthBench Professional、HLE、BrowseComp 和 EHR-Complex 四项基准上超过 GPT-6 Astra、Claude Opus 5.5 和 Gemini 4 Argon。其中 HealthBench Professional 测的是真实医生提交给聊天机器人的临床请求,EHR-Complex 则测智能体查询医院病历数据库的能力。

Body

训练方法上,蚂蚁从预训练阶段就把医疗数据与通用数据混合,再把医疗推理、工具调用、搜索和医生风格的对话行为拆成独立技能,分别用 RL 奖励在经医生和校验者核对过的任务上训练,最后通过多教师 on-policy 蒸馏把这些专精能力合并进单一模型。

MIT 开源加上专门针对电子病历查询等临床场景的测试设计,意味着医院和医疗科技公司可以直接拿权重做私有化部署和二次开发,这对医疗 Agent 赛道的通用大模型是一个直接挑战。

Original · rohanpaul_ai

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.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include LingSante, Ant Group, medical AI, MoE.

Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).