Chinese LLaMA-Alpaca 2

Chinese LLaMA-Alpaca 2 is the second phase of Yiming Cui’s influential open-source project bringing Meta’s LLaMA-2 models to Chinese: extended Chinese vocabulary, continued pretraining and instruction-tuned Alpaca variants, including 64K long-context models. It has around 7,100 stars on GitHub.

Overview

The original LLaMA models tokenize Chinese inefficiently and understand it poorly. This project fixes both problems: it expands the tokenizer with a large Chinese vocabulary so Chinese text is encoded far more efficiently, continues pretraining on Chinese corpora, and releases instruction-tuned Alpaca-2 counterparts trained for chat and task following. The lineup spans several sizes and includes long-context versions supporting 64K tokens, trained with techniques such as FlashAttention-2 and NTK-aware scaling. Complete training and inference recipes are published so others can reproduce or extend the work, and the models became a standard baseline for Chinese open-source LLM development. The project is maintained in 2026 and released under the Apache-2.0 license (subject to the LLaMA-2 community license for the base weights).

Key Features

• Chinese-extended LLaMA-2 models with a much larger Chinese vocabulary
• Instruction-tuned Alpaca-2 variants for chat and tasks
• 64K long-context models for long-document work
• More efficient Chinese tokenization than the original LLaMA-2
• Full training, fine-tuning and inference recipes included
• A standard reference for Chinese open-source LLM research

💳 Pricing

Chinese LLaMA-Alpaca 2 is free and open source. The models and code are published on GitHub and Hugging Face at no cost, under the Apache-2.0 license and Meta’s LLaMA-2 terms for the underlying weights.

Who Is It For

Chinese LLaMA-Alpaca 2 is for Chinese-language NLP developers and researchers who need capable open models with strong Chinese understanding, and for anyone studying how to adapt English-centric LLMs to other languages.