Chinese-LlaMA2 adapts Meta’s Llama 2 models for the Chinese language, closing the gap left by the original release, whose training data and tokenizer gave it only limited Chinese ability. The project provides Chinese-adapted base models and supervised fine-tuned chat versions that are fully open and usable commercially under the Llama 2 license terms. It is hosted on GitHub, where it has earned roughly 730 stars.
When Llama 2 arrived in 2023, Chinese developers quickly found that its vocabulary and pretraining mix under-served Chinese text. Chinese-LlaMA2 responds with continued adaptation on Chinese corpora and instruction tuning with open datasets such as UltraChat and Chinese Alpaca-style instruction data, producing models like Chinese-LlaMA2-chat-7B that converse naturally in Chinese while retaining Llama 2’s general abilities. The fine-tuned weights are published on Hugging Face, and the repository documents the data recipes and training approach so others can reproduce or extend the adaptation for their own domains. It complements the better-known Chinese-LLaMA-Alpaca project with its own data mix and release cadence. For builders who needed a Chinese-capable open model in 2023, it offered a ready starting point instead of a months-long adaptation project.
Chinese-LlaMA2 is free and open source. There are no fees or paid tiers โ you download the adapted weights and run them on your own hardware.
Chinese-LlaMA2 is for developers and researchers building Chinese-language chatbots, assistants or fine-tunes who want an openly licensed Llama 2 derivative that actually handles Chinese well.