Baichuan restructures to focus on AI for healthcare

Background

Baichuan is an AI company that builds large language models, and part of its business had been aimed at enterprise customers — the "To B" side — including work for the financial industry ^[5]. Its founder, Wang Xiaochuan, had invested in medical technology companies including Dr. Chunyu, Airdoc and DeepCare ^[2]. That made healthcare a direction the company already had ties to when it decided to narrow its scope.

What happened

The change came in steps rather than in a single announcement. On 19 February 2025, the prompt-engineering (PE) team inside the To B business was moved under the R&D group that oversees algorithms ^[5]. On 3 March 2025, the To B team responsible for financial-industry business was disbanded, and its staff signed resignation agreements that day ^[5]. According to 36Kr, co-founder Ru Liyun told employees on 3 March 2025 that the company had to focus on medicine ^[5]. In March 2025 Baichuan announced that it was restructuring its team to concentrate on applying AI to healthcare ^[1].

Put together, the sequence reads as a deliberate retreat from one customer segment and a redirection of people and attention toward another. A prompt-engineering team was not cut but re-homed inside algorithm research; the financial-industry team was dissolved outright, with departures taking effect the same day the message was delivered to staff ^[5].

The medical models

Baichuan-M2-32B, which the company described as its second open medical model, was presented as the world's strongest open-source medical model ^[3]. It is built on Qwen2.5-32B and released under the Apache License 2.0, which allows both research and commercial use ^[3]. The company reported a score of 60.1 on HealthBench, ahead of the open model gpt-oss-120b at 57.6 ^[3]. According to its model page, the 4-bit quantised version can run on a single RTX 4090 ^[3] — a consumer graphics card rather than a data-centre cluster. In January 2026 the company released Baichuan-M3, described as a new-generation open medical-enhanced large language model ^[4].

Why it matters

Three things make this reorganisation worth recording. First, it is a plain example of a model company giving up a paying enterprise segment to bet on a single vertical: the financial-industry To B team was dissolved and its staff signed resignation agreements on the day the decision was communicated ^[5]. Second, the new direction matched the founder's earlier investment history in medical technology ^[2], so the pivot was less a sudden idea than a return to a field the company's leadership already knew.

Third, the models themselves were released openly. An Apache 2.0 licence on a 32B model means other teams can take the weights and build on them, for research or for commercial products, without negotiating access first ^[3]. The claim that a 4-bit version runs on one RTX 4090 sets a low hardware bar ^[3], which matters for hospitals, clinics and small developers who do not own large GPU clusters. Running a model locally also keeps patient-related data inside an organisation's own machines, a practical concern in medical settings.

The medical direction also changes what counts as evidence. The comparison the company published rests on a benchmark score rather than on clinical use ^[3], and a benchmark result is only one kind of evidence about how a model behaves. The later release of Baichuan-M3 in January 2026 shows the focus on medicine outlasted the initial reorganisation ^[4].