Huawei and Zhejiang University Unveil DeepSeek-R1-Safe AI Model

At the recent Huawei Global Connect Conference, Huawei Technologies and Zhejiang University jointly introduced DeepSeek-R1-Safe, a groundbreaking foundation model built on Huawei's Ascend 1000 computing platform. This collaboration marks a significant step forward in addressing critical challenges at the intersection of AI performance and security.

A New Standard for AI Safety

Professor Ren Kui, Dean of Zhejiang University's School of Computer Science and Technology, detailed the model's innovative framework. "DeepSeek-R1-Safe represents a comprehensive approach to secure AI development," he explained. The model incorporates:

• A high-quality secure training corpus
• Balanced optimization techniques for security training
• Proprietary software/hardware integration

The framework specifically targets fundamental security challenges in large-scale AI training processes.

Unprecedented Security Performance

Test results demonstrate exceptional capabilities:

• 100% defense rate across 14 categories of harmful content (toxic speech, political sensitivity, illegal activity incitement)
• Over 40% success rate against jailbreak attempts
• 83% comprehensive security score, outperforming comparable models by 8-15%

Remarkably, these security gains come with minimal performance trade-offs. In standard benchmarks (MMLU, GSM8K, CEVAL), the model shows less than 1% performance loss compared to non-secure counterparts.

Industry Implications and Open Access

Zhang Dixuan, President of Huawei's Ascend Computing Business, emphasized the company's commitment to collaborative innovation: "By open-sourcing this technology through ModelZoo, GitCode, GitHub and Gitee, we're enabling broader participation in secure AI development."

The release signals growing industry recognition of security as a foundational requirement rather than an afterthought in AI systems.

Key Points:

• First domestic foundation model on Ascend 1000 platform
• Achieves security-performance balance through novel framework
• Outperforms competitors by significant margins
• Now available through major open-source platforms