Baichuan open-sources the Baichuan-M2 medical model

Background

Medical AI models are built to handle health and clinical questions, and this release was the open publication of one such model by Baichuan. ^[4][1]

What happened

TMTPost reported on 11 August 2025 that Baichuan had unveiled Baichuan-M2, an open-source medical model for private deployment in clinical settings. ^[4] The open release was Baichuan-M2-32B, its second medical-enhanced model, which the company published under the Apache 2.0 licence. ^[1]

The model builds on Qwen2.5-32B. ^[2] Baichuan called it the world's strongest open-source medical model. ^[1]

Key details

On HealthBench the model scored 60.1 overall and 34.7 on HealthBench-Hard. These results were ahead of gpt-oss-120b and DeepSeek-R1-0528 on the same benchmark, according to the company. ^[1] The company's figures are its own measurements.

It can run on a single RTX 4090 when quantised to 4 bits. ^[2] Quantisation stores the model's numbers with fewer bits, which reduces the memory needed to run it.

The technical report was posted to arXiv on 2 September 2025. ^[3]

Some terms in this entry need explaining. HealthBench is a benchmark for health conversations, and the figures here include its overall score and two subsets, Hard and Consensus. Private deployment means a model is run on an organisation's own servers rather than through an outside service, which matters where patient data is involved. Apache 2.0 is a permissive open-source licence that allows use, modification and redistribution, subject to its conditions.

A medical benchmark score is meaningful only alongside the test design, the grading criteria and the model settings used, so the technical report is an important companion to any figure in this entry.

Impact

The release added to the set of openly available medical models in this archive, with a stated purpose of research and reference. Because the scores are the company's own reports, readers should check the technical report for test conditions.