Chinese tech giant Huawei has co-developed a safety-focused version of artificial intelligence model DeepSeek that it said is "nearly 100% successful" in preventing discussion of politically sensitive topics.
Chinese regulators have required domestic AI models and the applications they power to reflect China's "socialist values" before they are released to the public, in compliance with tight controls on speech, Reuters reported.
Huawei said in a publication on a company WeChat account late on Thursday that it used 1,000 of its Ascend AI chips to train the large-language model, which was tweaked from DeepSeek's open-source model R1.
Huawei's partner was the elite Zhejiang University, the alma mater of DeepSeek's founder, Liang Wenfeng. DeepSeek and Liang, however, had no apparent involvement in the project. DeepSeek did not immediately respond to a request for comment.
CHINA INC EMBRACES, TWEAKS DEEPSEEK
DeepSeek's release of DeepSeek-R1 and V3 shocked Silicon Valley and tech investors outside China due to their level of advancement, triggering a selloff of Western AI stocks in January.
The AI models have been embraced, modified, and deployed across Chinese industry and society.
Chinese AI chatbots like Baidu's Ernie Bot - China's first answer to OpenAI's ChatGPT - refuse to answer or engage with many questions about Chinese domestic politics or topics considered sensitive by the ruling Communist Party.
Huawei's tweaked model is called DeepSeek-R1-Safe. Testing showed it to be "nearly 100% successful" in defending against "common harmful issues ... including toxic and harmful speech, politically sensitive content, and incitement to illegal activities," the company said.
That success rate dropped to 40%, however, when the behaviours were disguised by scenario-based challenges, role-playing scenarios, and encrypted coding, according to Huawei.
"Its comprehensive security defence capability reached 83%, outperforming multiple concurrent models like Qwen-235B and DeepSeek-R1-671B by 8% to 15% under identical testing conditions," the company added, referring to a model developed by Chinese tech giant Alibaba.
DeepSeek-R1-Safe exhibited a less than 1% performance degradation compared to the original DeepSeek-R1, Huawei said.
The company is holding its annual Huawei Connect conference in Shanghai, where on Thursday it broke years of secrecy about its chipmaking efforts to announce chip and computing power product roadmaps.
