LingSante 3.1 flash released
- Source
- huggingface.co
- Author
- InclusionAI model releases
- Date

Context
LingSante-3.1-flash is the Ling team's new health and medicine model, and the most visible change from its predecessor is the base model: it is built on the much larger general model Ling-3.1-flash. The team, part of Ant Group, published the weights under the MIT license through its inclusionAI organization on Hugging Face. According to the model card, the model keeps Ling-3.1-flash's architecture, with 560 billion parameters in total, 25 billion active for each token and a of up to 1 million tokens. The earlier Ling-3.0-flash-Sante had 124 billion parameters with 5.1 billion active.
It is a model. Each expert layer holds 512 small expert networks, and a router sends every token through 8 of them plus one shared expert, so a token touches only a fraction of the model. The whole model still has to be loaded to serve it, and the card's quickstart runs it on eight Nvidia H200 GPUs, or four of the newer GB300s.
The card says the medical training runs through every stage. Medical and general data were mixed during , mid-training and supervised fine-tuning, with the general data meant to keep skills such as coding, search and . Reinforcement learning then targeted medical reasoning, tool use, search and professional behavior, and a final distillation step, in which the model learns from several teacher models, combined those strengths into one model. Ant's stated aim is a model that can check evidence across sources, give professional judgments and complete multi-stage medical workflows, rather than only answer questions.
The performance claims are Ant's own, and the card's text gives no scores. It says the model delivers leading performance among open-source models on complex medical reasoning and ranks second overall on benchmarks of medical professional behavior, behind only GPT-6 Astra; the results appear only in a chart image that was not read for this summary. The card also limits how the model may be used. It is meant for research and evaluation, is not a medical device, and must not independently diagnose, prescribe, decide treatment or make triage decisions. The card adds that benchmark and simulated-environment results do not establish clinical effectiveness.
- Context and context window — Context is the information available for a generation; the context window is the capacity available to hold it.
- mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
- pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
- tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
Checking sign-in…
Loading comments…



