SHANGHAI/BEIJING: Chinese artificial intelligence startup DeepSeek released the first update to its hit R1 reasoning model in the early hours of Thursday, stepping up competition with US rivals such as OpenAI.
DeepSeek said via developer platform Hugging Face that R1-0528 was a minor version upgrade of R1 that nevertheless significantly improved its depth of reasoning and inference capabilities, including better handling of complex tasks, bringing its performance closer to OpenAI's o3 reasoning models and Google's Gemini 2.5 Pro.
The launch of R1 in January went globally viral, sent tech shares outside China plummeting, and challenged the view that scaling AI requires vast computing power and investment. Since R1's release, Chinese tech giants like Alibaba and Tencent have released models claiming to surpass DeepSeek's.
Thursday's update was initially light on details in contrast to the launch of R1 in January which was accompanied by a multi-authored academic paper that the AI community worldwide has parsed to understand the firm's strategies.
The Hangzhou-based firm said later in a short post on X that R1-0528 featured improved performance. In a longer post on WeChat, DeepSeek said the rate of "hallucinations", false or misleading output, was reduced by about 45-50% in scenarios such as rewriting and summarizing.
It said the update also enabled it to creatively write essays, novels and other genres, and had improved capabilities in areas such as generating front-end code and role-playing.
"The model has demonstrated outstanding performance across various benchmark evaluations, including mathematics, programming, and general logic," DeepSeek said.
DeepSeek's success has upended beliefs that US export controls were holding back China's AI advancements, after it released AI models that were on a par or better than industry-leading models in the United States at a fraction of the cost.
The startup added on Thursday that a variant of its update was created by taking the reasoning process used by the R1-0528 model, to then further enhance Chinese tech giant Alibaba's Qwen 3 8B Base model, a process known as distillation. The result was a performance surpassing the original Qwen 3 model by over 10%.
"We believe that the chain-of-thought from DeepSeek-R1-0528 will hold significant importance for both academic research on reasoning models and industrial development focused on small-scale models," DeepSeek added.
Bloomberg reported the update on Wednesday. It said that a DeepSeek representative had told a WeChat group it had completed what it described as a "minor trial upgrade" and that users could start testing it.
In response to competition from Deepseek, Google's Gemini has introduced discounted tiers of access while OpenAI cut prices and released an o3 Mini model that relies on less computing power.
Deepseek is still widely expected to release R2, a successor to R1. Reuters reported in March, citing sources, that R2's release was initially planned for May. DeepSeek also released an upgrade to its V3 large language model in March.
&uuid=(email))