A hospital in a tier-three city cannot ship patient records to a distant cloud. Its doctors still want a model that reasons about symptoms. That gap, between capable frontier systems and deployable local software, is exactly where Baichuan (百川智能) planted its flag in August 2025.
The company did not try to beat GPT-class models on every benchmark. It released a model that does one hard thing well, and that fits on a single consumer GPU.
Key takeaways
- Baichuan-M2, an open-source medical-enhanced large model (医疗增强大模型), was released on 11 August 2025.
- Its 32B version scored 60.1 on HealthBench (vendor-stated), a result Baichuan says exceeds OpenAI's open-weight gpt-oss-120B at 57.6.
- Quantized and compressed, the model can run on a single NVIDIA RTX 4090, which Baichuan estimates cuts deployment cost to about 1/57 of a DeepSeek-R1 H20 dual-node setup (vendor claim).
- It is the follow-up to Baichuan-M1 (14B, January 2025), described by the company as the first medical-enhanced open model.
- The release reflects a deliberate strategic pivot: from a general-purpose contender toward medical AI as the core business.
From generalist to specialist
Baichuan was founded by Wang Xiaochuan (王小川), the former CEO of Sogou, and spent its early years chasing the general large model (大模型) race with Baichuan-7B, Baichuan-13B and the Baichuan4 series. By early 2025 the company had narrowed its focus to healthcare. Baichuan-M1, a 14B model released in January 2025, was positioned as the first medical-enhanced open-weight release. Seven months later, Baichuan-M2 extended that line.
The throughline is a tool Baichuan calls the "AI patient simulator" (AI患者模拟器). The company says it used real data to construct tens of thousands of synthetic patients of varying age, sex and symptom profiles, then simulated millions of consultation episodes. Baichuan states that training on medical interaction data, combined with end-to-end reinforcement learning, pushed the model's clinical reasoning further than the earlier M1 release.
The benchmark claims, as stated
On the HealthBench evaluation, Baichuan reports Baichuan-M2 scores 60.1. On HealthBench Hard, a stricter subset, it reports 34.7. Both figures are vendor-stated and have not been independently reproduced here. Baichuan frames 34.7 as making the model the second release worldwide (after OpenAI's GPT-5) to clear 32 on that subset; that framing is the company's own and should be read as a marketing claim, not an audited result.
The 32B size is the headline: Baichuan says this smaller open model outranks OpenAI's open-weight gpt-oss-120B (57.6) on HealthBench. A 32B model beating a 120B open model on a domain test is plausible through specialization, but readers should treat the comparison as vendor-reported until third-party HealthBench runs confirm it.
Why a single RTX 4090 matters
The more durable story is deployment economics. Baichuan states that after quantization the model runs on one RTX 4090 with near-lossless precision, and that this costs roughly 1/57 of a DeepSeek-R1 H20 dual-node deployment. The company also says its MTP (multi-token prediction) variant lifts single-user token throughput by 74.9%. These are vendor claims; the underlying point, that medical buyers need private, on-premise inference, is sound regardless of the exact multiplier.
For a clinic, "open weight" is not a philosophical position. It is the difference between keeping records inside the building and sending them out. Baichuan says it adapted the model for mainstream domestic Chinese chips as well, which would let hospitals use existing hardware.
The broader line: M3
Baichuan has also listed Baichuan-M3 on its site as an open, commercially usable model with about 235 billion parameters and a HealthBench score reported at 65.1. M3 is positioned as a stronger reasoning model with low hallucination and "full consultation" capability. Treat the 65.1 figure as vendor-stated. The existence of M1, M2 and M3 in under a year shows a steady cadence in one vertical rather than a scatter of general releases.
Why medicine, and why now
Baichuan's pivot to healthcare is best read against the wider Chinese model market. General-purpose labs burned 2023 and 2024 racing on broad benchmarks that few buyers could operationalize, while hospitals and insurers sat on a concrete, regulated need that off-the-shelf models did not meet. By specializing, a smaller firm trades headline generality for a defensible niche where deployment constraints—privacy, on-premise hosting, domain accuracy—matter more than raw scale. That is a calmer, more durable path than chasing every frontier release.
Honest limitations
- Every benchmark number above (HealthBench 60.1, HealthBench Hard 34.7, the gpt-oss-120B comparison, 1/57 cost, 74.9% throughput) comes from Baichuan's own announcements and has not been independently reproduced in this article.
- "Medical-enhanced" does not mean clinically validated or regulator-approved for diagnosis. The model is a reasoning assistant, not a certified medical device.
- Performance on general, non-medical tasks was not the focus of Baichuan's claims; do not assume M2 matches general-purpose models outside healthcare.
- Open-weight licensing terms and any commercial restrictions should be read directly from the model card before deployment.
What readers can do now
- If you evaluate models for a clinic or insurer, benchmark Baichuan-M2 yourself on your own de-identified cases rather than trusting vendor HealthBench numbers.
- Read the official model card and license on Baichuan's repository before assuming free commercial use.
- Watch whether independent HealthBench runs reproduce the 60.1 score; that is the real test of the open-medical-model thesis.