DeepSeek R1-0528 is slightly behind OpenAI's o4 mini and o3 reasoning models in terms of code generation, but ahead of xAI's Grok 3 mini and Alibaba's Qwen 3.
Chinese AI startup DeepSeek announced on Wednesday in its official community group that the DeepSeek R1 reasoning model has completed a minor version trial upgrade.
DeepSeek released R1-0528 on the developer platform Hugging Face but did not publish a description or comparison results for the model.
On the LiveCodeBench benchmark ranking developed by researchers from the University of California, Berkeley, MIT, and Cornell University, DeepSeek's updated R1 model performs slightly below OpenAI's o4 mini and o3 inference models in code generation and outperforms xAI's Grok 3 mini and Alibaba's Qwen 3.
Earlier this year, DeepSeek released the R1 model, which is comparable to or even outperforms leading industry models in the US, yet at a cost that is only a fraction of the latter, sparking widespread discussion.
In March, DeepSeek released a minor version upgrade of its V3 large language model, improving performance on reasoning-related tasks and achieving scores exceeding GPT-4.5 on math and code-related evaluation datasets.
