PaddleNLP is Baidu’s large language model and NLP library built on the PaddlePaddle framework. It packs a large model zoo and a complete LLM toolkit — training, alignment, compression and serving — into one library, and with about 13,000 stars on GitHub it is among the most-used open NLP toolkits in the Chinese AI ecosystem.
PaddleNLP aims to make the whole LLM lifecycle a few lines of code. Its model zoo covers the ERNIE family alongside popular open architectures such as LLaMA and Qwen, with unified loading and generation APIs. The LLM toolkit spans pre-training, supervised fine-tuning and preference alignment with methods such as DPO, plus quantization and high-performance inference, all wired for distributed training on PaddlePaddle. Beyond chat models, the library is known for its applied NLP strengths: the UIE universal information extraction model, sentiment analysis, semantic and neural search, document intelligence and question answering, many exposed through one-line Taskflow-style pipelines. For teams standardized on the Paddle ecosystem — or working heavily with Chinese language data — PaddleNLP is the natural counterpart to what Hugging Face libraries provide elsewhere, and it has stayed in active development through 2026.
PaddleNLP is free and open source under the Apache-2.0 license. Everything runs on your own infrastructure; there are no paid tiers for the library itself.
PaddleNLP is for NLP and LLM engineers who want a complete training-to-deployment toolkit on the PaddlePaddle stack, particularly for Chinese and bilingual applications such as information extraction, search and enterprise document processing.