As the new semester begins, Shenzhen University has officially rolled out the full-scale version of DeepSeek-R1 (671B), a fully China-developed and locally deployed AI model, in the first week of the academic term. Developed by the National Engineering Laboratory for Big Data System Computing Technology at Shenzhen University and the Shenzhen University Information Center, in collaboration with Huawei, the full-scale DeepSeek-R1 (671B) model is now running on the Ascend platform, offering efficient and stable AI services for the entire university community.
Compared to the distilled version, the full-scale one boasts significant improvements. With a massive 671 billion parameters, it delivers enhanced performance, enabling complex logical reasoning and in-depth interactions with long texts. It also offers a wider range of capabilities, including the integration of private knowledge bases, providing tailored services for teaching and research.
The National Engineering Laboratory for Big Data System Computing Technology has long focused on research related to large models and their applications. In April 2023, it open-sourced Linly,China's first large-scale Chinese language model with 10 billion parameters, based on the LLaMA framework (https://github.com/CVI-SZU/Linly). In collaboration with Tencent, the Laboratory also launched TencentPretrain, the first modular framework for training large language models(https://github.com/Tencent/TencentPretrain), significantly advancing the development of large language model technology in China. Building on the Linly model, the team has also released specialized models for various vertical domains,including interactive 2D/3D digital humans, virtual customer service agents, facial granularity analysis multimodal models, as well as diagnostic models for pathology, ultrasound, and abdominal CT scans, along with models for financial software requirements analysis.Moving forward, the team will continue to enhance system performance based on DeepSeek.
