llamachinese/llama-chinese

Llama-Chinese, run by a group calling itself the Llama Chinese Community, is a hub for working with Meta's open-source Llama language models in Chinese. Llama is the family of large language models that Meta has released for free, including Llama 2, Llama 3, the smaller Llama 3.2 mobile-friendly models, and the most recent Llama 4 multimodal mixture-of-experts release. The project's stated goal is to build the best Chinese open-source ecosystem around these models, with everything released for commercial use. A lot of the value sits in the models the community has trained and shared. The main one is Atom, a Chinese pre-trained model built on top of Llama, available in 1B, 7B, and 13B parameter sizes through Hugging Face, ModelScope, and WiseModel. There are also Chinese fine-tuned versions of Llama 2 and Llama 3 published by the community, and pointers to the official Meta models. The README links to Atom-7B-Chat as their flagship conversational model and mentions newer pre-training runs using 2.7 terabytes of Chinese text. For people who want to use these models, the README walks through several setup paths. You can install it with Anaconda, run it inside Docker, use the lightweight llama.cpp runner, launch a Gradio web interface, build an API service, or run the models through ollama. Beyond just running the models, the project includes scripts for further pre-training, LoRA fine-tuning, and full-parameter fine-tuning, plus instructions for loading the resulting checkpoints back in. The repository also covers production concerns. There are notes on quantising models to smaller sizes, on speeding up inference with TensorRT-LLM, vLLM, JittorLLMs, and lmdeploy, and on extending the model through LangChain. A section on benchmarks compares the Chinese performance of Llama 2, Llama 3, and Llama 4 against each other. Beyond the code, the community offers shared compute resources, training data, a forum, an app store at llama.family, and online events for members.