China's DeepSeek Upgrades R1 AI Model, Narrowing Gap with Western Counterparts The latest version of the R1 model reportedly performs just below OpenAI's o3 and o4-mini, based on evaluations by LiveCodeBench

Chinese AI startup DeepSeek has quietly released an upgraded version of its R1 model, 'DeepSeek-R1-0528', on the open-source platform Hugging Face, claiming substantial improvements in mathematical reasoning, programming, and logic, alongside a reduction in hallucination rates.

The update positions DeepSeek's model closer to top-tier systems such as OpenAI's o3 and Google's Gemini 2.5 Pro, according to performance data cited on Hugging Face. The company's low-key release strategy continues its trend of disrupting the AI landscape without formal fanfare.

"In terms of comprehensive performance, the updated model is approaching the level of the industry's leading systems," DeepSeek said in its Hugging Face release.

DeepSeek first drew global attention in January with the debut of its original R1 model, which delivered competitive performance despite limited resources. The company's progress raised concerns in the West, with R1's launch contributing to brief market volatility and a dip in Nvidia's stock price.

The latest version of the R1 model reportedly performs just below OpenAI's o3 and o4-mini, based on evaluations by LiveCodeBench, an AI benchmarking platform. The model's capabilities are built on cost-efficiency and reasoning-focused design—traits that have made DeepSeek a symbol of China's broader AI ambitions.

Founded in Hangzhou, DeepSeek has become a key player in China's push to rival American AI giants, navigating restrictions on advanced semiconductors by focusing on optimisation and resourceful development. The company's founder, Liang Wenfeng, has gained domestic prominence, recently participating in a high-level economic forum chaired by President Xi Jinping. His rise reflects growing confidence in China's home-grown AI talent, amid intensifying global competition in the sector.