Baichuan releases the open Baichuan-M3 medical model

Background

Baichuan-M3 is the company's new-generation open-source medical-enhanced LLM, and it was presented at a media briefing. ^[1][2]

What happened

On 13 January 2026 Baichuan released Baichuan-M3, its new-generation open-source medical-enhanced LLM, at a media briefing. ^[1] The model has 235 billion parameters, builds on Qwen3-235B-A22B and is licensed under Apache 2.0. ^[2]

Baichuan said the model surpassed GPT-5.2 on HealthBench, HealthBench-Hard and hallucination evaluation. ^[2] Its scores on those tests are the company's own claims. The model scored 44.4 on HealthBench-Hard, 28 points above Baichuan-M2. ^[2]

Key details

Training introduced SPAR, a staged clinical-workflow reward method, and a Fact-Aware RL method. ^[2] SPAR rewards each stage of a clinical workflow separately, while Fact-Aware RL checks medical claims against evidence during reinforcement learning.

The technical report followed on arXiv in February 2026. ^[3]

Several terms here come from the technical field. A hallucination is a confident statement that a model makes without support in fact. HealthBench-Hard is a harder subset of the HealthBench test, so scores on it are usually lower than on the full set. A technical report is the document in which developers describe a model's design, training and results.

Medical AI models are often judged on whether they can handle a case step by step, from taking a history to proposing a diagnosis, because clinical work does not happen in a single answer. Readers who want to test such claims should look for the evaluation settings, the list of tests used and any independent replication.

Impact

The release continued the company's open medical line, with a model available under an open licence. Because the figures are reported by the company, readers should check the technical report and the benchmark definitions before relying on them. Readers should also keep the technical report and the company's own release wording side by side when comparing results.