DeepSeek improves R1 - now AI answers more accurately and makes fewer mistakes
DeepSeek has released an updated version of its R1 artificial intelligence model, DeepSeek-R1-0528. This new version is faster, smarter, and less likely to make mistakes called "hallucinations" in AI - when the model comes up with incorrect or false answers.
Here's What We Know
DeepSeek has long attracted attention because their R1 model is among the top 10 best large-scale language models, and it trains faster and cheaper than products from such giants as OpenAI or Google.
The updated R1 was tested on various challenging tests: the AIME 2024 and 2025 maths competitions, Google-Proof QA tests, coding, and even one of the most difficult tests, Humanity's Last Exam. Although the model has improved, it still only answers 17% of questions correctly on the most difficult exam. The situation is the same for leading competitors: complex tasks are still very difficult for AI.
The main changes in the new version are a significant reduction in errors and better accuracy of answers. Now DeepSeek-R1-0528 is less likely to mislead.
For those who want to work with the model on their own, DeepSeek has made an open source version with 8 billion parameters that can be run on a powerful Nvidia 4090 graphics card. This makes the technology more accessible to developers and researchers.
It is worth noting that China uses DeepSeek artificial intelligence to create modern combat aircraft. DeepSeek provides new engineering approaches, speeds up the design cycle, and helps focus on key technical challenges.
Source: DeepSeek
