At the China Judicial Artificial Intelligence Conference (CJAI2026) held recently in Shanghai, the Tsinghua University Internet Law Research Institute launched a large-scale open-source legal model called LegalOne-R1. The development of this model was strongly supported by Huawei Cloud's Ascend AI cloud service, marking another significant step forward in the application of artificial intelligence in the legal field in China.

LegalOne-R1 offers three different parameter versions: 1.7B, 4B, and 8B, aiming to provide powerful intelligent support for legal work through deep training on Chinese judicial data. The model adopts a "middle training + post-training" dual-stage enhancement strategy, integrating instruction fine-tuning and reinforcement learning, which can effectively inject legal knowledge and simulate professional workflows, thereby achieving natural emergence of legal thinking.

In practical applications, LegalOne-R1 not only has the ability to remember provisions and analyze concepts but can also perform multi-hop reasoning and analysis of adjudication logic, making its performance in real business scenarios more stable, accurate, and practical. In addition, the model's training was supported by the computing power from Huawei's AI Hundred-University Program, as well as full collaboration from the National Key Laboratory of Internet System Architecture and other industry partners.

According to public evaluations, LegalOne-R1-8B demonstrates excellent legal professional capabilities. It outperforms many larger parameter general models in multiple legal evaluation sets such as LexEval, LawBench, and JecQA, showing a leading level in core tasks such as legal concept understanding, law provision memorization, and multi-hop reasoning.

The full-process training and inference adaptation of this model were completed based on the Huawei Ascend Atlas910B computing platform and the MindSpore AI framework. To better serve judicial institutions and legal technology companies, the parameters of LegalOne-R1 were publicly released on domestic and international open-source platforms on January 23. In the future, technical reports and application guides will also be released to help more users understand and apply this cutting-edge technology.