Baichuan-M3

What it is

Baichuan-M3 is a new-generation open-source medical-enhanced language model from Baichuan, the Chinese AI company. It was released at a media briefing on 13 January 2026. ^[1] It has 235 billion parameters, the numerical settings a model learns in training, and it builds on Qwen3-235B-A22B. ^[2]

The model is open-source and released under Apache 2.0. ^[2]

Origins

Baichuan-M3 follows Baichuan-M2 and is described as a new generation of the company's open medical line. ^[1][2]

Key features

Baichuan reported that M3 scored 44.4 on HealthBench-Hard, a harder subset of the HealthBench medical evaluation. That is 28 points higher than Baichuan-M2, and the model ranks first on the HealthBench Total leaderboard. ^[2]

Baichuan said M3 surpasses GPT-5.2 on HealthBench, HealthBench-Hard, a hallucination evaluation and its own SCAN-bench clinical benchmark. ^[2] The technical report says the model achieves state-of-the-art results on HealthBench and on the new HealthBench-Hallu and ScanBench tests, and that it significantly outperforms GPT-5.2 in clinical inquiry, advice and safety. ^[3] On SCAN-bench, M3 ranks first in clinical inquiry, laboratory testing and diagnosis, leading the next model by 12.4 points in clinical inquiry. ^[2]

Training method

M3 was trained with SPAR, which stands for Step-Penalized Advantage with Relative baseline. SPAR splits a clinical workflow into history taking, differential diagnosis, lab testing and final diagnosis, and gives each stage its own reward. ^[2] A reward is the score a model receives for a good step, and staged rewards let training check each part of the workflow separately.

A second method, Fact-Aware RL, checks medical claims against authoritative evidence online during reinforcement learning. The aim is to reduce hallucinations, which are confident statements that are not supported by facts. ^[2]

Deployment

Four-bit weight quantisation reduces the model's memory use to 26% of the original. FP8 and GPTQ-INT4 variants are also available. ^[2] Quantisation stores the model's numbers with fewer bits, so the same model needs less memory to run.

Position in development

The technical report, arXiv:2602.06570, was posted in February 2026. ^[3] The scores come from Baichuan's own reports and benchmarks, so readers should consult the report for the test conditions.