Baichuan launches Baichuan-M1-preview and open-sources Baichuan-M1-14B

Background

The January 2025 launch included two models: Baichuan-M1-preview, a reasoning model, and the open medical model Baichuan-M1-14B. ^[4][1]

What happened

On 24 January 2025 Baichuan released Baichuan-M1-preview, which it called China's first all-scenario deep-thinking model. The model reasons across language, vision and search. ^[4] The company said the preview outperformed o1-preview on maths and code evaluations. ^[4] It also offered an evidence-based medical mode intended to support doctors in decisions. ^[4]

At the same time Baichuan open-sourced Baichuan-M1-14B, a small version of M1-preview, in Base and Instruct versions. ^[5][1]

Key details

Baichuan described Baichuan-M1-14B as the industry's first open-source large language model trained from scratch for medical scenarios. ^[1] The model was trained on 20 trillion tokens of medical and general data. ^[2] A token is a small piece of text, and the figure describes the amount of text used in training.

On Baichuan's medical benchmarks the model averaged 72.23, close to Claude 3.5 Sonnet's 74.85. ^[2] These are the company's own published results, measured on its own medical test suite.

The technical report was posted to arXiv on 18 February 2025. ^[3]

Two terms matter here. A reasoning or deep-thinking model works through a problem in steps before it answers, which usually takes more computation than a direct reply. o1-preview and o1-mini are reasoning models from OpenAI, and they are the reference points in the comparisons Baichuan made. Decision-support tools are designed to inform a professional's judgement rather than make decisions on their own.

Impact

The release gave researchers and developers an openly available model built for medical work, along with a technical report and a set of published results to check against. Its figures are the company's own reports, so anyone using the model in a clinical setting should consult the report and the licence terms. Readers can compare it with the later medical models in this archive.