Baichuan

★ 3/5 · Open Source Model

Baichuan Intelligent Technology is a Beijing lab founded by former Sogou CEO Wang Xiaochuan that has narrowed from general-purpose Chinese LLMs to medical AI, publishing the Apache-2.0 licensed Baichuan-M3-235B and Baichuan-M2-32B clinical models on Hugging Face alongside a consumer AI-doctor app. Its earlier Baichuan2 and Baichuan4 general models remain available, but the company's current work and benchmarks are almost entirely healthcare.

Pros

Genuinely open weights under Apache 2.0 with research and commercial use both permitted, and the flagship Baichuan-M3-235B posts 44.4 on HealthBench-Hard and 65.1 on HealthBench overall while ranking first across all three SCAN-bench clinical dimensions, beating GPT-5.2 on the lab's own published comparisons
the release ships as FP8, GPTQ-INT4 and Q4_K_M GGUF variants where W4 quantisation cuts memory to 26% of the original and Gated Eagle3 speculative decoding adds a 96% speedup, while the smaller Baichuan-M2-32B fits on a single RTX 4090 and every variant serves through vLLM or SGLang behind an OpenAI-compatible endpoint.

Cons

The company walked away from the general-purpose race, so this is not a Qwen or DeepSeek substitute for broad work
the corporate site and the API console are Chinese-only with no English toggle, and self-serve signup assumes a mainland account, while enterprise and medical API access now routes through a business-enquiry form rather than a checkout
no inference provider currently serves Baichuan-M3 on Hugging Face, leaving you to self-host 235B parameters or a quantised cut of it
the general Baichuan4 line looks stale next to the M-series, and the model card itself warns the output is for research and reference only and cannot replace professional diagnosis.

Cost

Open weights are free under Apache 2.0 for both research and commercial use, which is the only route that works without a Chinese account. The mainland API console publishes Baichuan4-Air at 0.98 li per 1,000 tokens, roughly 0.98 yuan per million tokens, and states Baichuan4-Turbo is priced at 80% of GPT-4o; there is no USD rate card, no self-serve international billing, and the medical M-series API is quoted through a business-enquiry form rather than listed.

Verdict

Best for teams building Chinese-language medical or health assistants who can self-host a large open model and want a permissive licence; look elsewhere if you need a general-purpose hosted API with an English console and published USD pricing.

Visit Baichuan