DeepSeek R1 Gets a Minor Update with Major Upgrades, Performance Closely Approaches OpenAI o3!

Yesterday, DeepSeek quietly released the latest version of its R1 large model—DeepSeek-R1-0528, which is now available for public beta testing. Always keeping a low profile, DeepSeek did not attach detailed technical explanations to this release. Instead, it simply informed users in its official WeChat community that “the DeepSeek R1 model has completed a minor version trial upgrade,” and invited everyone to test it on the official website, app, and mini-program.

Hugging Face URL: https://huggingface.co/deepseek-ai/DeepSeek-R1-0528

Judging from user feedback, this so-called “minor update” still brings substantial improvements, particularly in reasoning and output capabilities. Specifically, the new DeepSeek R1:

• Enhanced Reasoning Ability: The model demonstrates more structured and logical performance in “Chain-of-Thought” reasoning.
• Improved Text Output Quality: Generated text is more organized, with a more mature linguistic style close to leading models like Google’s. –
• Optimized Code Generation: In programming tasks, R1’s output is more coherent, and the code quality is higher.

Reddit user B89983ikei said after using it: “It’s really amazing at programming!! In just one conversation, it created a complete game with multiple options and fully playable!! One significant change I noticed in the new R1 is… it’s even stronger in programming!! But it failed some (unknown) deductive reasoning challenges… ones it used to answer correctly!! Another obvious change is that it now reasons differently and thinks in the user’s native language instead of always defaulting to English.”

However, B89983ikei added: “For me, I enjoy solving ‘abstract’ reasoning problems… so I feel this version of R1 is weaker in that aspect. I guess it’s because its thinking style has changed… it focuses more on precise problems and neglects more abstract ones. I’m not sure if this is a good thing!! For me, it’s a step backward… When we remove those more abstract thinking elements, maybe we’re losing something closer to ‘real thinking’!! If you know what I mean… Has anyone else noticed this!?” “Also, I don’t like the new text formatting… it’s more confusing! Not as comfortable as before… too many unnecessary emojis!! These are constructive criticisms from a user who likes and has been using DeepSeek!!”

At the same time, some users noticed a slight increase in response time for the new DeepSeek R1, but most believe this is a worthwhile trade-off for more accurate results. After testing, users also provided the first benchmark for the new DeepSeek R1!

The new Deepseek R1-0528 performs nearly on par with o3 (high) in the LiveCodeBench benchmark.

Netizen Yuchen Jin on X also found through testing: “My vibe check: It seems to be the only model that can always correctly answer ‘What is 9.9 – 9.11?’.”

When comparing with the old version, OpenRouter revealed: “The new DeepSeek R1 has been trained on 100 million tokens and is still growing, including a free version!” The comparison chart is as follows:

Overall, since its establishment in 2023, DeepSeek has been focused on building high-performance general artificial intelligence models. Models like DeepSeek v3 and R1 have achieved excellent results in multiple evaluations. Notably, R1 is open-source under the MIT license, allowing free use, modification, and even commercial projects—greatly contributing to the popularization and innovation of AI technology.

Have you tried the latest R1 model? What do you think? Feel free to leave a comment and share your experience!