China’s DeepSeek releases update to AI model that sent US shares tumbling earlier this year

In a bold move to bolster its competitive edge, Chinese AI startup DeepSeek unveiled an update to its R1 reasoning model early Thursday. This new version, dubbed R1-0528, was introduced on the developer platform Hugging Face, although an official announcement from the company is still pending. While specifics about the model and comparisons to others were not disclosed, the LiveCodeBench leaderboard—developed by researchers from esteemed institutions such as UC Berkeley, MIT, and Cornell—places DeepSeek’s updated model just shy of OpenAI’s o4 mini and o3 reasoning models in code generation capabilities, while outperforming xAI’s Grok 3 mini and Alibaba’s Qwen 3. Reports from Bloomberg indicate that a DeepSeek representative shared in a WeChat group that this update is considered a “minor trial upgrade,” and users are encouraged to begin testing it. Earlier this year, DeepSeek challenged the narrative that U.S. export controls were stifling China’s AI progress by releasing models that not only matched but occasionally surpassed American counterparts, all at significantly lower costs. The introduction of the original R1 in January sent tech stocks outside of China into a downward spiral, contradicting the notion that advanced AI development necessitates substantial computing resources and investment. Following the R1’s launch, major Chinese tech firms, including Alibaba and Tencent, have claimed to develop models that outstrip DeepSeek's capabilities. In response, Google’s Gemini has rolled out more affordable access options, while OpenAI has reduced prices and launched an o3 mini model that requires less computing power. Anticipation builds for the release of R2, DeepSeek’s next iteration, which was initially rumored to debut in May, according to Reuters sources. Additionally, the company upgraded its V3 large language model back in March.