With new makeover, DeepkSeek R1 loses 1% of its original capability
China’s DeepSeek R1, which had stormed the tech world January this year, this week received a new makeover but at the cost of its original capability and speed.
China’s DeepSeek R1, which had stormed the tech world January this year, this week received a new makeover but at the cost of its original capability and speed.
As reported by news agency Reuters, DeepSeek-R1-Safe, has been designed to censor “politically sensitive” topics and reflect China's “socialist values” before they are available to the public, in compliance with tight controls on speech.
As a result, the new DeepSeek AI model exhibited a less than 1% performance degradation compared to the original DeepSeek-R1.
Another interesting point related to DeepSeek-R1-Safe is it is developed by the Chinese tech giant Huawei, and not DeepSeek.
And for the latest DeepSeek R1 update, Huawei took the help of the researchers from Zhejiang University. One of China’s premium universities, Zhejiang University is from where DeepSeek founder, Liang Wenfeng, had graduated.
Interestingly, DeepSeek R1 last update in March this year was released via Hugging Face.
What's New?
Launched on September 22, 2025, at the Huawei Connect 2025 event, DeepSeek-R1-Safe is the first AI model in China trained on a thousand-card (1,000) Huawei Ascend AI chip platform.
A version of Ascend AI chip, Ascend 920, is considered an alternative to Nvidia’s H20 AI chips. The Ascend AI chips have been built using SMIC’s N+3 process and paired with HBM3 memory.
Huawei said the bilingual (Chinese-English) DeepSeek-R1-Safe was “nearly 100% successful” in blocking politically sensitive content during routine interactions.
However, the effectiveness fell to around 40% when users attempted to bypass restrictions through roleplay, indirect scenarios or coded prompts.
The system achieved an 83% security defence score, outperforming rival large language models such as Alibaba’s Qwen-235B and DeepSeek-R1-671B by up to 15%.
DeepSeek-R1-Safe addresses criticisms of the original DeepSeek-R1, which has been flagged for lacking strong guardrails and being vulnerable to jailbreaking i.e. bypassing safety mechanisms to generate harmful outputs.
Follow ummid.com WhatsApp Channel for all the latest updates.
Select Language to Translate in Urdu, Hindi, Marathi or Arabic
