Deepseek Releases Update To Its AI Reasoning Model R1

The AI startup DeepSeek released the latest update to its reasoning AI model, R1, on the Hugging Face platform on Thursday. The update, R1-0528, was launched without an official announcement and with only a few details provided.

According to TechCrunch, the Chinese company shared a post on the social media platform WeChat on Wednesday, informing followers about the recent development. The update includes a “minor” upgrade: the adoption of a permissive MIT license, which allows the model to be used commercially.

“In the latest update, DeepSeek R1 has significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimization mechanisms during post-training,” states the document shared by DeekSeek on Hugging Face. “Its overall performance is now approaching that of leading models, such as o3 and Gemini 2.5 Pro.”

In the graphics of the benchmark test results, DeepSeek R1-0528 performs just as good—or better—than similar competitive models. In the AIME 2025 benchmark, DeepSeek R1-0528 reached an 87.5 score, below the OpenAI o3 model with 88.9 points, but better than Gemini-2.5 Pro 0506, Qwen3-235B, and its own previous version, DeepSeek-R1.

“Beyond its improved reasoning capabilities, this version also offers a reduced hallucination rate, enhanced support for function calling, and a better experience for vibe coding,” states the document.

Clément Delangue, co-founder and CEO of Hugging Face, spread the news of the R1 update on the social media platform X.. “Just a few minutes later & the updated R1 is already available on some of our inference partners,” wrote Delangue. “All on the model page – beautiful!” Multiple users shared their interest in the latest version of R1.

DeepSeek was recently involved in a data breach case, in which its database was exposed, offering access to third parties to around 1 million logs, API keys, and chat history. The vulnerability has already been handled by DeepSeek.