Chinese Llama 2 7B, from LinkSoul AI, was the open-source community’s first downloadable and runnable Chinese-adapted LLaMA-2 model. Starred around 2,200 times on GitHub, it extended Meta’s LLaMA-2 7B with a much larger Chinese vocabulary and Chinese pretraining so the model could actually read and write Chinese efficiently.
The original LLaMA-2 tokenizer handles Chinese poorly, splitting characters into byte fragments. This project expands the vocabulary with Chinese tokens, then continues pretraining on Chinese text and releases both base and instruction-tuned chat versions. The result tokenizes Chinese far more efficiently and follows Chinese instructions โ a recipe many later Chinese open models built upon.
The practical effect of the vocabulary expansion is easy to measure: Chinese text that LLaMA-2 would encode as many byte-level tokens is represented with far fewer tokens, which shortens sequences, speeds up generation, and leaves more of the context window for actual content. The project also documented its training data sources and released training scripts, so it served as a template for many subsequent Chinese adaptations of open base models.
Deployment options documented by the project include standard Transformers inference and quantized variants for smaller GPUs, and the instruction-tuned version can hold multi-turn Chinese conversations. While today’s Chinese open models are far stronger, this release remains a clean, well-documented example of the vocabulary-extension-plus-continued-pretraining recipe, which is why it is still referenced in courses and surveys.
The project is free and open source under the Apache-2.0 license; the base model remains subject to Meta’s LLaMA-2 community license terms.
Researchers and developers interested in Chinese LLM adaptation, tokenizer extension techniques, or the history of open Chinese language models.