Baichuan 2
Open multilingual language models with 7B and 13B parameters.
Baichuan 2 is a series of open large language models released by Baichuan Intelligence in September 2023. It includes 7 billion and 13 billion parameter versions trained on 2.6 trillion tokens. The models support Chinese and English tasks and allow free commercial use with a licence.
- Trained on 2.6 trillion tokens of data.
- Supports a 192K token context window.
- Free commercial use with official licence.
- Releases intermediate training checkpoints.
- Available in 7B and 13B parameter sizes.
What it is
Baichuan 2 is a family of large language models created by Baichuan Intelligence. The series launched in September 2023 with two sizes: seven billion and thirteen billion parameters. Both base and chat variants are available. The company trained the systems on 2.6 trillion tokens of high quality data. It claims strong results on Chinese and English benchmarks. The models are designed for general purpose text generation and reasoning tasks.
How it came about
Baichuan Intelligence announced the model in September 2023 via a technical report. The team released base and chat versions for both parameter counts. In December 2023, they updated the thirteen billion chat model to version two. This update aimed to improve mathematics, logic, and instruction following. The developers also shared eleven intermediate training checkpoints for research purposes.
Inside the model
The architecture uses a decoder only transformer design. Training relied on 2.6 trillion tokens from diverse sources. The company states it ensured data compliance during this process. For faster inference, the models utilise scaled dot product attention from PyTorch 2.0. A four bit quantised version is provided for the chat models to reduce memory usage. The system supports a context window of up to 192,000 tokens.
What it can do
The models perform well on standard tests like MMLU and C-Eval. The thirteen billion base model scored 58.1 on C-Eval and 59.17 on MMLU. It outperforms similar sized open models such as LLaMA2-13B. The company highlights strength in vertical domains including law, medicine, and code. Benchmark results show competitive performance in multilingual translation and general knowledge.
How to use it
Weights are openly available under the Apache 2.0 licence. Commercial use is permitted after obtaining a free licence via email. The company does not provide its own applications on mobile or web platforms. Developers can run the models locally using PyTorch 2.0. Intel processors and GPUs are supported through specific optimisation libraries.
Safety and testing
The developers warn against using the model for illegal activities or harming social security. They advise against deploying it in internet services without proper safety reviews. The company disclaims liability for issues arising from misuse, data leaks, or public opinion risks. They state that efforts were made to ensure training data compliance.
Why it matters
The release of intermediate checkpoints helps researchers study training dynamics. By offering free commercial licences, the project encourages broader adoption in industry. It provides a strong baseline for Chinese language processing in the open source community.
Questions people ask about Baichuan 2
Can I use Baichuan 2 for commercial products?
Yes. You may use it commercially after requesting and receiving an official licence from Baichuan Intelligence via email. The weights are open under Apache 2.0 terms.
What hardware do I need to run it?
You need a PyTorch 2.0 environment. The company recommends using Intel Core or Xeon processors with BigDL-LLM for better inference performance. Four bit quantised versions are available to save memory.
How many parameters does the largest model have?
The largest version in the Baichuan 2 series has thirteen billion parameters. There is also a seven billion parameter variant available for both base and chat tasks.
Good reading and sources
There is no English Wikipedia article on Baichuan 2 yet, so the facts on this page come from the sources listed above: huggingface.co/baichuan-inc/Baichuan2-13B-Chat, arxiv.org/abs/2309.10305. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.