Recently, Baichuan Intelligence officially released its open-source medical-enhanced large model Baichuan-M2, which has attracted widespread attention in the industry. Just five days ago, OpenAI released two new models and claimed that their medical capabilities were the strongest and the deployment cost was extremely low.
However, Baichuan-M2 quickly surpassed others with its smaller size and outstanding medical capabilities, becoming the top open-source model globally.
Baichuan-M2 scored 60.1 on the HealthBench evaluation, surpassing OpenAI's latest released gpt-oss120b model (scored 57.6), and leading other open-source large models such as Qwen3-235B, Deepseek R1, Kimi K2 internationally. Notably, Baichuan-M2's design philosophy highly emphasizes the privacy needs of medical users, allowing for private deployment, fully meeting the requirements of various medical institutions.
In terms of technology, Baichuan Intelligence has optimized Baichuan-M2 to an extreme level of lightweight processing. The model's precision after quantization is nearly lossless and can be deployed on a single RTX4090 card. This deployment method reduces costs by about 57 times compared to the DeepSeek-R1H20 dual-node deployment. This advantage allows many medical institutions to quickly deploy and apply the model with existing hardware conditions.
To improve application effectiveness in scenarios requiring rapid interaction such as emergency care and outpatient services, Baichuan Intelligence also launched the Baichuan-M2MTP version optimized based on the Eagle-3 architecture. This version achieved a 74.9% increase in token processing speed in single-user scenarios.
With stronger medical capabilities, will Baichuan-M2 affect its general capabilities? Baichuan Intelligence's research shows that high-quality medical data plays an important role in improving the model's general performance. Therefore, Baichuan-M2 not only did not decrease in core abilities such as mathematics, instruction following, and writing, but actually improved, making it have broad application potential beyond the medical field.
Additionally, Baichuan-M2's ability to handle complex medical issues is comparable to GPT-5. When GPT-5 was released, OpenAI specifically emphasized that this model was the only one worldwide to score over 32 points in the HealthBench Hard evaluation. Baichuan-M2 scored 34.7 points, becoming the second model worldwide to exceed 32 points, demonstrating its strong strength in the medical field.
Project: https://huggingface.co/baichuan-inc/Baichuan-M2-32B
Key Points:
1. 🌟 The Baichuan-M2 model released by Baichuan Intelligence leads globally in medical capabilities, scoring 60.1 points.
2. 💡 The model has been lightweighted, allowing deployment on a single card, significantly reducing costs for medical institutions.
3. 🚀 Baichuan-M2's ability to handle complex medical problems is comparable to GPT-5, showing great application potential.