- Huawei announced DeepSeek-R1-Safe, a safety-focused AI version designed to block politically sensitive content and malicious language. The project was developed in conjunction with Zhejiang University, where DeepSeek founder Liang Wenfeng once studied, but Huawei asserts that Mr. Liang and DeepSeek were not directly involved.
- The system was trained on 1,000 Huawei-produced Ascend chips, based on the open-source DeepSeek-R1 platform but adjusted to comply with Beijing’s regulations, ensuring the AI reflects “socialist values” and adheres to online speech controls.
- In tests, DeepSeek-R1-Safe achieved nearly 100% in preventing responses related to politically sensitive topics, malicious language, and illegal activities. However, in complex scenarios such as roleplay, challenging scripts, or coding, effectiveness dropped to 40%.
- Huawei reported the system achieved an overall security defense score of 83%, 8–15% higher than Alibaba Qwen-235B and DeepSeek-R1-671B under the same testing conditions. Notably, additional safety measures reduced performance by less than 1% compared to the original version.
- The launch comes as China intensifies its adoption of DeepSeek technology. After R1 and V3 were released in early 2025, Western markets were rattled, causing a sell-off of AI stocks in Silicon Valley.
- Many domestic AI platforms like Baidu Ernie Bot already restrict responses on political topics, and DeepSeek-R1-Safe is seen as a next step to institutionalize these limitations in advanced AI systems.
- The launch coincides with the annual Huawei Connect conference in Shanghai, where the company also unveiled new chip roadmaps and computing infrastructure, revealing its semiconductor ambitions after years of secrecy.
📌 With DeepSeek-R1-Safe, Huawei is positioning politically-oriented safe AI, achieving an 83% defense score and surpassing Alibaba Qwen-235B by 8–15%. In tests, DeepSeek-R1-Safe achieved nearly 100% in preventing responses related to politically sensitive topics, malicious language, and illegal activities. However, in complex scenarios such as role-playing, challenging scripts, or coding, effectiveness dropped to 40%. The system was trained on 1,000 Huawei Ascend chips, indicating an ambition to consolidate China’s AI and semiconductor position.
