Baichuan 2 is a series of open large language models from Baichuan Intelligence, released in 7B and 13B sizes for both base and chat use. Trained on 2.6 trillion tokens with a focus on Chinese and English, the models set strong benchmark results for their size on MMLU, C-Eval and GSM8K when released. The project has about 4,100 stars on GitHub.
The series emphasizes training quality over parameter count: a carefully filtered multilingual corpus, RoPE positional encoding with context up to 4096 tokens, and alignment through SFT and RLHF for the chat versions. Baichuan 2’s technical report details its data and scaling-law methodology, making the release as much a training reference as a model drop. Both sizes ship on the Hugging Face Hub with Transformers support, quantization examples (INT8/INT4) and fine-tuning scripts in this repository, and the chat models handle bilingual dialogue, knowledge questions and code. Baichuan 2 became one of the most-adopted open Chinese LLM families, spawning many community fine-tunes for legal, medical and financial domains, and it remains a dependable self-hosted option where bilingual capability matters.
The release was notable for its transparency about data: Baichuan described sourcing from books, papers, code and web data with aggressive quality filtering, and published intermediate checkpoint evaluations so researchers could study how capabilities emerge with scale. The chat models support system prompts and function-style interactions used by Chinese enterprise deployments, and the 13B model in particular was a common base for domestic industry fine-tunes. Both sizes run comfortably on a single high-memory GPU when quantized.
Baichuan 2 is free: the code is Apache-2.0 and the weights are released under the Baichuan 2 model license, which allows research and commercial use subject to its terms.
Baichuan 2 is for developers and teams who need a proven open bilingual (Chinese–English) language model to fine-tune or self-host for chat, knowledge and domain applications.