Huawei Launches DeepSeek-R1-Safe AI System Focused on Safety and Compliance

Huawei has rolled out a revised version of its DeepSeek artificial intelligence system, called DeepSeek-R1-Safe. This new model emphasises safety and adherence to regulations, reflecting the increasing governmental oversight in China regarding advanced AI tools.

Overview of DeepSeek-R1-Safe

The DeepSeek-R1-Safe model was created in partnership with Zhejiang University, a prominent academic institution in China and the alma mater of DeepSeek’s founder, Liang Wenfeng. However, Huawei confirmed that neither Liang nor the DeepSeek company had direct involvement in this initiative.

System Specifications and Features

Based on the open-source DeepSeek-R1 framework, the system has been retrained utilising 1,000 of Huawei’s Ascend AI chips. The primary objective was to integrate measures that prevent the model from participating in politically sensitive discussions, producing harmful speech or incentivising illegal behaviours.

Effectiveness and Performance

Huawei asserts that DeepSeek-R1-Safe achieved “nearly 100% success” in preventing politically sensitive content during standard interactions. However, its effectiveness dropped to around 40% when users attempted to navigate around restrictions using roleplay, indirect scenarios, or coded language. Overall, the model performed with an 83% security defence score, surpassing competing large language models, including Alibaba’s Qwen-235B and DeepSeek-R1-671B, by as much as 15%. Notably, Huawei indicated that these safety enhancements minimally impacted the model’s performance, reducing it by less than 1% compared to the original version.

Alignment with Local Priorities

This release demonstrates Beijing’s commitment to ensuring AI platforms embody “socialist values” while adhering to stringent limits on online expression. Other domestic platforms, such as Baidu’s Ernie Bot, are already blocking responses concerning politically sensitive topics. Huawei’s enhancement seems to formalise similar controls within advanced AI systems.

Global Trends in AI Development

The advancement also mirrors a larger global trend of adapting AI to align with local priorities. For instance, in early 2025, the Saudi Arabian firm Humain introduced an Arabic-native chatbot with a focus on Islamic culture and values. Such initiatives highlight how countries are making efforts to shape AI not merely for technical prowess but also to reflect cultural and political priorities.

Launch Event and Future Plans

Huawei revealed the model at its annual Connect conference in Shanghai, where it also shared its long-term plans for semiconductors and computing infrastructure. This announcement comes at a time of increasing adoption of DeepSeek technology within China, alongside ongoing global discussions regarding the balance of AI innovation, safety, and regulation.