At the recently concluded Huawei Global Connect Conference, Huawei Technologies Co., Ltd. jointly launched with Zhejiang University the first domestic foundation model based on the Ascend 1000 computing platform - DeepSeek-R1-Safe. This innovative product aims to address current security and performance issues in the AI field, opening a new chapter in intelligent technology.

Professor Ren Kui, Dean of the School of Computer Science and Technology at Zhejiang University, introduced the core innovations of this model in detail. DeepSeek-R1-Safe was built through a full-process secure post-training framework, which includes a high-quality secure corpus, balanced optimization for secure training, and self-innovative software and hardware platforms. The design of this framework aims to solve key issues in secure training at the fundamental level.

Notably, DeepSeek-R1-Safe has achieved breakthroughs in secure training with trillions of parameters. Its defense capabilities are impressive. Test data show that the model achieves an overall defense success rate close to 100% in 14 dimensions of harmful information, such as dealing with toxic and harmful speech, politically sensitive content, and incitement to illegal activities. In terms of defense against various jailbreak modes, the success rate also exceeds 40%. The comprehensive security defense capability is as high as 83%, outperforming similar models Qwen-235B and DeepSeek-R1-671B by 8% to 15%.

Additionally, in general ability benchmark tests such as MMLU, GSM8K, and CEVAL, the performance loss of DeepSeek-R1-Safe is controlled within 1%, indicating that it not only enhances security protection but also ensures the usability of the model, successfully achieving a balance between security and performance.

Zhang Dixuan, President of Huawei's Ascend Computing Business, stated at the conference that Huawei is actively promoting basic software innovation and AI security capabilities. Through open collaboration with universities and industry partners, Huawei is driving technological advancement. Meanwhile, the model has been fully open-sourced in communities such as ModelZoo, GitCode, GitHub, and Gitee, allowing more developers and researchers to participate.

This milestone release not only brings new hope to the AI security field but also paves the way for the coordinated development of the future AI industry ecosystem.