one year on
DeepSeek releases R1-0528 reasoning model, matching top-tier benchmarks
The MIT-licensed update posts scores competitive with OpenAI o3 and Gemini 2.5 Pro on maths and coding benchmarks, while adding JSON output and function calling.
DeepSeek quietly released R1-0528, an updated version of its open-weight reasoning model, on Hugging Face today under the MIT license. The company’s API docs highlight improved benchmark performance, enhanced front-end capabilities, reduced hallucinations, and new support for JSON output and function calling.
Early benchmark comparisons show R1-0528 scoring near OpenAI o3 and Gemini 2.5 Pro on standard maths and coding evaluations, making it one of the strongest open-weight reasoning models available. The update is available immediately on Hugging Face and via the DeepSeek chat interface.
The release has sparked discussion in AI communities about how a ‘minor update’ from a Chinese lab can match models that were considered frontier just months ago. Some researchers noted the continued MIT licensing as a significant advantage for downstream development.
The record
Called the update a 'stealth release' that puts pressure on US labs with its open-weight approach.
One year later — open only if you can handle spoilers
R1-0528 became a widely used baseline for reasoning model comparisons throughout mid-2025, especially among open-weight enthusiasts. DeepSeek continued to iterate with further releases later in the year.