Baichuan AI

📊 SOTA sweep: HealthBench 65.1 / Hard 44.4 🥇 Hallucination 3.5% (lower than ChatGPT) 🛡️✨ ScanBench all-stations #1: 74.9 / 72.1 / 74.4 🏆

✨Key takeaways: SPAR: align RL credit to where decisions happen — optimize stage-wise, not via one noisy end reward. Fact-Aware RL: verify atomic claims with retrieval → make hallucination measurable & optimizable🧐 Rubric Evolution: auto-mine & patch adversarial reward hacks.

Baichuan-M3 Technical Report is live. 🚀 Report：arxiv.org/abs/2602.06570 Models: hf.co/collections/baichuan-i… Try: ying.ai Built for clinical decision support, not trivia QA, optimized on the real outpatient workflow: Inquiry → Lab Testing → Diagnosis 🧑⚕️→🔬→🧾

We have officially open-sourced the new-generation medical LLM Baichuan-M3, which scored 65.1 on HealthBench and claimed the championship with 44.4 on HealthBench Hard. It has comprehensively outperformed GPT-5.2 across the board in the medical field. @OpenAI @claudeai

Evaluation shows that the medical hallucination rate of M2 Plus is significantly lower than that of general LLMs, about three times lower than DeepSeek. Now it’s available on web (ying.ai) and in mobile app stores.

We have just launched our enhanced evidence-based medical model, Baichuan-M2 Plus, while simultaneously upgrading its supporting application Baixiaoying and opening the API.

Today we introduced Baichuan-M2, a medically-enhanced reasoning model, which surpasses all open-source models including gpt-oss-120b on the #HealthBench Benchmark. Try it here: huggingface.co/baichuan-inc/… #AIHealth

In terms of visual, speech, and multi-modal streaming processing, Baichuan-Omni-1.5 supasses GPT-4o mini, and its leading edge is even more pronounced in the field of multi-modal medical applications.

We launched Baichuan-Omni-1.5 open-source omni-modal large model. It not only supports omni-modal understanding of text, image, audio, and video, but also has dual modalities generation capabilities for both text and audio.

It has unlocked a medical evidence-based model, enabling a complete end-to-end service from medical evidence retrieval to deep reasoning. To better promote AI healthcare ecosystem, we also launched the industry's first open-source medical enhancement large model, Baichuan-M1-14B.

We released China's first full-scenario deep thinking model, Baichuan-M1-preview. It’s currently the only one in China that simultaneously possesses reasoning capabilities in 3 areas: language, vision, and search. It has surpassed o1-preview in multiple authoritative evaluations.

Its performance in financial scenarios surpasses GPT-4 by nearly 20%! platform.baichuan-ai.com/fin…

We officially release the full-link domain-enhanced large model Baichuan4-Finance. Built on high-quality financial data, it achieves simultaneous improvement in financial and general capabilities through an industry-first domain self-constrained training scheme.

The solution includes a full-chain high-quality general training data, two models (Baichuan4-Turbo & Baichuan4-Air), and a comprehensive domain enhancement toolchain. The "1+3 Product Matrix" achieves a multi-scenario availability rate of up to 96%, a leading figure of industry.

Baichuan AI launched the one-stop commercial solution aimed at helping enterprises achieve private deployment of LLMs at the lowest cost and best efficiency. platform.baichuan-ai.com/

Developer Experience Center: platform.baichuan-ai.com/pla…

We also have announced the opening of 4 APIs: Baichuan 4, Baichuan3-Turbo, Baichuan3-Turbo-128K and Assistant API. All developers and enterprise users can log in Baichuan Developer Center to conveniently and efficiently experience powerful capabilities of new generation models.

Baixiaoying is an AI assistant who knows how to search, stands apart from local peers, due to its ability to “search like a professional” while “guiding and inspiring users to ask accurate questions”. Now it’s available on web (ying.ai) and in mobile app stores.

Recently we released the latest generation LLM, Baichuan 4, and launched our first AI assistant "Bai Xiao Ying". Baichuan 4 has greatly improved in all capabilities, what makes the model's abilities rank first in the authoritative domestic model evaluation published by SuperCLUE.

Furthermore, Baichuan 3 has made broken through in "iterative reinforcement learning" technology, further enhancing the semantic understanding and generation abilities. It excels in aspects such as format, rhythm, and expressive meaning in poetry creation, surpassing other LLMs.