(Yicai) May 29 -- DeepSeek-R1-0528, the newly released updated version of the large language model of Chinese artificial intelligence startup DeepSeek, has coding capabilities nearly matching those of OpenAI o3, ChatGPT developer's most powerful reasoning model.

DeepSeek released the upgrade on open-source community Hugging Face earlier today. Despite the model card and official announcement having not yet been updated, developer tests on code testing platform LiveCodeBench revealed its improved capabilities.

The DeepSeek Assistant notified users in its official community group that DeepSeek-R1 had undergone a minor version trial upgrade late yesterday. However, developers pointed out that this "minor upgrade" was far from insignificant.

In addition to DeepSeek-R1-0528's upgraded coding skills, developers highlighted other enhancements, including improved performance in writing tasks, a more natural language style, and better formatting. However, it is still unclear how the new model compares to Anthropic's Claude Opus 4, which is considered the world's best coding model.

Earlier in March, DeepSeek similarly released the DeepSeek-V3-0324 on Hugging Face, which adopted reinforcement learning techniques from R1's training process, significantly boosting performance on reasoning tasks and optimizing capabilities in front-end development and Chinese writing.

DeepSeek has not yet announced the highly anticipated R2, with overseas users commenting that their expectations for R2 have significantly grown after the latest R1 update. Beyond the updated model capabilities, developers hope the DeepSeek-R2 will substantially extend context length and add multimodal support, which are critical for production use.

Editor: Martin Kadiev