Huawei has launched a safety-centred version of the DeepSeek artificial intelligence model, claiming it is almost entirely effective at preventing politically sensitive discussions.
Huawei unveils safety-focused DeepSeek AI
The model, dubbed DeepSeek-R1-Safe, was developed in partnership with Zhejiang University, a leading institution and the alma mater of DeepSeek’s founder Liang Wenfeng. Huawei clarified that neither Liang nor DeepSeek were directly involved in the project.
According to a post on Huawei’s official WeChat account, the company trained the system using 1,000 of its Ascend AI chips. The project builds on the open-source DeepSeek-R1 model, but adapts it to comply with Beijing’s regulatory requirements that AI products reflect “socialist values” and align with strict controls on online speech.
Performance and testing
Huawei said the model was “nearly 100 per cent successful” in preventing responses involving politically sensitive issues, toxic language or illegal activity prompts. However, the success rate dropped to 40 per cent in more complex tests that disguised prompts through roleplay, scenario-based challenges or coded inputs.
Overall, Huawei reported the system achieved an 83 per cent comprehensive security defence score, outperforming rival large-language models such as Alibaba’s Qwen-235B and DeepSeek-R1-671B by 8–15 per cent under the same conditions. Crucially, the company added, the additional safety measures resulted in less than a per cent drop in performance compared with the original model.
Global impact of DeepSeek technology
The announcement comes as Chinese firms increasingly adopt and adapt DeepSeek’s technology, which drew global attention earlier this year after the release of its R1 and V3 models. Their sophistication rattled Western markets, sparking a sharp sell-off of AI-related stocks in Silicon Valley in January.
Domestic AI platforms, including Baidu’s Ernie Bot, already restrict responses to politically charged or sensitive topics, mirroring the government’s content red lines. Huawei’s DeepSeek-R1-Safe appears to be the latest step in formalising such controls within advanced AI systems.
The reveal coincided with Huawei’s annual Connect conference in Shanghai, where the company also lifted the lid on its semiconductor ambitions, unveiling new roadmaps for chips and computing infrastructure after years of closely guarded secrecy.
About the Author
