Huawei and Zhejiang University Unveil “DeepSeek-R1-Safe”

Huawei, together with Zhejiang University, introduced DeepSeek-R1-Safe. This is a safer version of the widely used DeepSeek AI model.

The system is designed to block politically sensitive and harmful content while keeping its strong reasoning abilities intact.

Key Features

• The model was built using 1,000 Huawei Ascend AI chips.
• It achieved nearly 100% success in avoiding politically sensitive discussions during normal interactions.
• Huawei says the model lost only 1% of its performance and speed compared to the original DeepSeek-R1.
• It provides stronger safety guardrails while still delivering accurate reasoning results.

Performance and Testing Results

• In testing, DeepSeek-R1-Safe reached an 83% overall security defense rate.
• This result is 8–15% better than Alibaba’s Qwen-2.5 model under the same conditions.
• However, the model’s success rate drops to only 40% when users try advanced methods such as role-playing, scenario challenges, or coded language.

Even with these limitations, the system still performs better than most competing AI systems in China.

Global Trend of National AI

• China is not the only country building AI systems that reflect national identity.
• Saudi Arabia recently launched an Arabic AI chatbot that reflects Islamic culture, values, and heritage.
• Other countries are also developing AI systems designed to reflect local values rather than relying only on Western models.
• Experts call this growing trend “AI nationalism”, where every country builds AI aligned with its own culture and regulations.

Open Source and Research Transparency

• DeepSeek has gained international attention for being highly open-source.
• The company has shared detailed research papers, including training methods and formulas.
• This openness has allowed researchers around the world to validate, adapt, and improve the model.
• Other groups have also built safer versions of DeepSeek, such as RealSafe-R1, which used a dataset of 15,000 safety examples.
• These projects show that AI can be trained to balance reasoning performance with stronger safety protections.

Security Risks and Vulnerabilities

• Despite progress, DeepSeek models still face major security risks.
• Independent analysis by Protect AI rated DeepSeek-R1 as “High Risk” for enterprise use.
• The models are vulnerable to jailbreaking, prompt injection, adversarial attacks, and misuse in business settings.
• DeepSeek openly acknowledges these issues. In fact, it became the first Chinese AI company to publish detailed safety risk data in the Nature journal.
• This step followed practices from American AI companies like OpenAI and Anthropic.

China’s Rules for Safer AI

• Chinese regulators require that all public-facing AI systems follow “socialist values” before release.
• This means AI must avoid topics that are considered politically sensitive or harmful.
• Similar restrictions already exist in other Chinese systems such as Baidu’s Ernie Bot, often referred to as China’s version of ChatGPT.
• This reflects Beijing’s strategy of ensuring technology growth always supports national priorities.

Global AI Race

• The global AI race has grown more intense in 2025.
• Worldwide, $200 billion was invested in AI infrastructure this year alone.
• The United States leads in computing power with nearly 20,000 megawatts of AI capacity.
• China leads in the number of large-scale data centers, with 230 facilities across the world.
• France has surprised many by ranking second globally in AI chips, with nearly one million units.
• Analysts say AI could add as much as $4.4 trillion in global productivity.

This is why governments see AI not as optional, but as strategic national investment.

News Gist

Huawei and Zhejiang University have launched DeepSeek-R1-Safe, a safer AI model designed to block politically sensitive and harmful content.

Released on September 19, 2025, it claims nearly 100% success in simple cases while keeping strong reasoning performance with minimal trade-offs.

FAQs

Q1. What is DeepSeek-R1-Safe?

It is a safer version of the DeepSeek R1 AI model, developed by Huawei and Zhejiang University to block sensitive or harmful content.

Q2. How effective is it?

In normal interactions, it avoided politically sensitive topics with nearly 100% success, though its effectiveness dropped to 40% under advanced jailbreak attempts.

Q3. How does it compare to rivals?

It outperformed Alibaba’s Qwen-2.5 model by 8–15% in safety tests, while maintaining almost the same reasoning ability as the original DeepSeek-R1.

Q4. Why is this important for China?

China requires AI systems to follow “socialist values”, making national alignment a must for public AI deployment.

Q5. Is DeepSeek-R1-Safe open-source?

Yes. The base model remains open-source, and Huawei applied its modifications for safety, offering transparency similar to previous DeepSeek projects.