Wink Pings

560B open-weight medical model beats GPT-6 Astra on three benchmarks

inclusionAI releases the open-weight medical mixture-of-experts model LingSante-3.1-flash, with 560B total parameters and 25B active, surpassing GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon on HealthBench Professional, HLE, and BrowseComp. From sleep analysis to interactive consultation, it treats medical tasks as workflows. The weights are open-sourced.

Open-weight medical models have taken another step forward. inclusionAI today released LingSante-3.1-flash, a mixture-of-experts (MoE) model with 560B total parameters and 25B active parameters, aimed at medical scenarios while retaining general capabilities.

For readers unfamiliar with MoE, here is one way to understand it: the total parameter count is large, but each inference activates only 25B of them, so cost and speed are close to a small model while capacity is close to a large model.

Official image:

官方配图

It is not intended to do medical question answering alone. The official positioning targets real medical problems: complex medical reasoning, long-horizon workflow execution, and professional behavior alignment. It scored 65.72 on HealthBench Professional, 72.22 on HLE (text-only, with tools), and 94.10 on BrowseComp (with context), surpassing GPT-6 Astra, Claude Opus 5.5, and Gemini 4 Argon on all three. For long-horizon medical workflows, it scored 57 on PhysicianBench and 74.40 on EHR-Complex.

In the official demos, from sleep analysis to evidence retrieval to interactive consultation, the model handles multi-step tasks:

One concrete scenario is sleep analysis. LingSante-3.1-flash calculates heart rate and HRV by sleep stage within the YASA framework, preserves heartbeat timestamps for signal review, and leaves fields blank when data is insufficient rather than forcing a conclusion.

Another is interactive consultation. Simulating a case of headache with unilateral weakness, the model recognizes red flags, updates recommendations as test results arrive, and finally organizes the entire conversation into a structured medical record.

Training methods also include some notable design choices. General data and medical data are mixed during pretraining, mid-training, and SFT, and multi-teacher on-policy distillation is used at the end to merge capabilities, improving medical ability while retaining general ability. Evidence-to-Capability Matching (ECM) aligns medical evidence with reasoning objectives; MEDPAR refines tasks through multi-agent review, and samples must pass medical correctness and task-consistency checks before entering training.

ECM

Multi-agent search uses contribution-aware rewards: compared with multi-agent RL that uses only accuracy as the reward, accuracy improves and total search steps drop by more than 20%.

搜索

Behavior alignment uses a behavior matrix to assign training rewards, with 99.62% consistency with physician annotations.

行为对齐

行为对齐2

Training data for long-horizon tasks covers real MCP, simulated MCP, and world-model environments; tasks evolve across scenarios, tools, and constraints, and results are confirmed with cross-trajectory checks and task-level verifiers.

长程任务

The weights are already open-sourced, and the model card shows the MIT license. The FP8 version is on HuggingFace and ModelScope, and an MXFP4 quantized version is also available. The API can run on Computrix, starting at $20 per month. For deployment, the model card provides launch commands for SGLang (recommended) and vLLM. The context window is 1 million tokens, and the architecture is Hybrid-linear MoE. The base model is Ling-3.1-flash; the previous medical version, Ling-3.0-flash-Sante, was 124B/5.1B, and this time it jumps directly to 560B, a sizable step.

Developer Paulo Steinberg said in the comments that open-weight models get stronger every week, and sooner or later someone will ask why we are still paying US labs; as for medical effectiveness, he leaves it to doctors to judge. Another netizen, ShadowAguy, joked that 560B parameters just to say go see a doctor. The model card also states clearly: for research and evaluation only, not a medical device, and cannot be used independently for diagnosis, prescription, or treatment. Results from benchmarks and simulated environments do not equal clinical effectiveness.

The significance of LingSante-3.1-flash is not another set of benchmark records. It pushes competition in medical models from 'answering questions' to 'doing work': sleep analysis, interactive consultation, and evidence retrieval are all real workflows. Whether the model can be deployed in clinical settings is another matter. At least the weights are out there, so you can run it yourself.

Model card · ModelScope · API

发布时间: 2026-11-23 01:36