Ailyre

Model release · Ant Group

Ant open-sources LingSante-3.1-flash, a medical MoE model with 560B total and 25B active parameters and a 1M context

Launched

Ant's Ling team has open-sourced LingSante-3.1-flash, a medical model built on Ling-3.1-flash's hybrid-linear MoE architecture, with 560B total parameters, 25B active per token and a 1M-token context window. It combines complex medical reasoning, long-horizon medical workflows, tool use, evidence retrieval and interactive consultation while keeping general skills. Ant says it leads the HBP medical-behavior benchmark (65.72) and trails only GPT-6 Astra on the three-benchmark medical-behavior average, and that contribution-aware multi-agent search training improves accuracy while cutting total search steps by over 20%. Weights are on Hugging Face and ModelScope under the MIT license, in the original FP8 version and an MXFP4 quantized version; SGLang is the recommended way to serve it, and vLLM is also supported. Ant says the model is for research and evaluation only, is not a medical device and must not be used as an independent basis for diagnosis, prescriptions or treatment.

Availability

Open weights under the MIT license on Hugging Face and ModelScope, in the original FP8 and an MXFP4 quantized version; also available via API on Computrix.

Source trace

Reviewed by Ailyre. Source evidence is presented separately from our summary.

Source excerpt

x.com

Source checked:

Source excerpt

x.com

Source checked: