Open-Source Medical Models Are Starting to Dare to Say 'I Don't Know'
A 560-billion-parameter open-source medical model, LingSante-3.1-flash, was released today. It outperforms GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon on three benchmarks: HealthBench Professional, HLE, and BrowseComp. More noteworthy is that when data is insufficient, it chooses to leave the answer blank rather than make one up.
LingSante-3.1-flash. The open-source medical model released today by Ant Ling.
560 billion total parameters, a MoE architecture that activates only 25 billion parameters per pass, and a 1 million token context window. The weights are public under the MIT license, and both FP8 and MXFP4 versions are available.
First, a detail. After reading the technical post, a user noted that when data is insufficient, this model chooses to leave it blank instead of inventing an answer. When new test results enter the conversation, it adjusts its earlier recommendations. When it encounters danger signals, it flags them. At the end of the conversation, it organizes everything into a structured medical record for a doctor to review.
This behavioral pattern is more noteworthy than the benchmark numbers. In medical settings, an AI confidently making things up is not a low-probability event. It is just that people usually do not call it out.
LingSante-3.1-flash outperforms GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon (the flagship models from three closed-source vendors) on three benchmarks:
- HealthBench Professional: 65.72
- HLE (text + tools): 72.22
- BrowseComp (with context): 94.10
Its evaluation is not limited to medical QA. PhysicianBench (57) and EHR-Complex (74.40) both test long-horizon clinical workflows: retrieving evidence, invoking tools, and staying coherent across multi-step tasks. This is the leap from 'answering questions' to 'executing tasks.'
Why do hospitals need open-source weights? Patient data cannot be sent to someone else's cloud. Privacy and compliance make on-premises deployment a hard requirement. The MXFP4 quantized version lowers some of the barriers.
Technically, the model inherits Ling-3.1-flash's hybrid linear MoE architecture, 63 KDA + 9 Gated MLA. The previous generation, Ling-3.0-flash-Sante, had 124 billion total parameters and activated 5.1 billion per pass; this time it is more than a fourfold increase. The training pipeline was medical mixed pretraining, midtraining, SFT, task-level reinforcement learning, plus multi-teacher on-policy distillation.
A few design choices developers should note:
ECM matches medical evidence to reasoning objectives. MEDPAR refines tasks through multi-agent review; before samples enter training, their medical correctness and task consistency are checked.
Contribution-aware multi-agent search improves accuracy and reduces search steps by more than 20% compared with RL rewarded purely on correctness.
Behavior alignment uses a behavior matrix as the training reward, with 99.62% agreement with physician annotations. On average medical behavior, it is second only to GPT-6 Astra.
Long-horizon task synthesis covers real MCP, simulated MCP, and world-model environments, with cross-trajectory checks and task-level verifiers confirming the results.
For deployment, a SGLang day-0 image is already supported, with MTP speculative decoding and Hopper optimizations built in. It can run on eight H200s (TP8+EP8). vLLM recipes cover H200, GB300, H800, H100, and B200. If you do not want to deploy it yourself, there is an API on Computrix; the Basic Token Plan costs $20 per month.
FP8 weights: Hugging Face | ModelScope
MXFP4 weights: Hugging Face | ModelScope
One developer commented, 'I'll have to beg the hospital to give me a machine.' The reply: 'You now have enough evidence to make your case.'
The model card states clearly: for research and evaluation only; it is not a medical device and must not be used independently for diagnosis, prescribing, or treatment decisions.
This is not a boilerplate disclaimer. A 560-billion-parameter model willing to admit when it does not know, and to stop where it should, is itself the kind of direction medical AI should take.
发布时间: 2026-11-23 00:35