As a pioneer in the AI field, DeepSeek once again sparked industry discussions with its technological breakthroughs. On May 28, 2025, DeepSeek quietly released version R1-0528. This update not only achieved significant performance improvements but also shocked the market with its free and open policy. This article, compiled by the editorial team of AIbase, delves into the highlights and potential of this version.
128K Super-Large Context Window, Enhanced Reasoning Ability
The upgrade to support a 128K super-large context window is one of the most eye-catching features of the DeepSeek R1-0528 version, providing broader space for handling complex tasks. Compared to previous versions, the new R1 performed excellently in text recall tests within a 32K context window, showing a significant improvement in accuracy, particularly suitable for scenarios requiring deep understanding and precise answers. However, tests showed a slight decline in performance at the 60K context level, and the results of the 120K test are still pending, with future performance worth looking forward to.
In addition, the new version has undergone deep optimization in reasoning capabilities. It is reported that R1-0528 can conduct deep reasoning on complex problems for 30 to 60 minutes, showcasing its "deep thought" characteristics. This "slow work produces fine results" reasoning style enables it to perform on par with OpenAI's o1 or even o3 level models in tasks such as mathematics, code generation, and natural language reasoning.
Significant Improvement in Code Generation and Writing Abilities
R1-0528's performance in code generation is particularly impressive. Latest tests show that its coding abilities have significantly improved, with both generation speed and accuracy enhanced. It is even considered comparable to the performance of OpenAI's latest model in Live CodeBench. Whether generating code snippets quickly or handling complex programming tasks, R1-0528 demonstrates strong competitiveness.
Meanwhile, the new version has undergone a comprehensive upgrade in writing tasks. Text generation is now more natural and fluent, with significantly improved formatting standards, suitable for various scenarios from academic writing to creative content. This improvement makes it more versatile in practical applications, offering users more efficient tools for creation.
Free Access, A Disruptor in the Industry?
Surprisingly, despite the significant upgrades, DeepSeek R1-0528 remains free to use. This strategy not only lowers the user threshold but also further consolidates DeepSeek's leading position in the open-source AI field. Compared to commercial models that require high subscription fees, R1-0528's "more features, same price" undoubtedly offers an attractive choice for developers, researchers, and ordinary users alike.
Industry insiders point out that DeepSeek's engineering strength is rapidly widening the gap with other large domestic models, and its inference speed and performance even make some international top-tier models feel pressured. The release of R1-0528 is not just a display of technology but also a challenge to the business model of the AI industry.
From this update, it is clear that DeepSeek is not satisfied with mere technological catching up. Instead, it aims to reshape the competitive landscape of the AI industry through continuous innovation and an open strategy. The open weights of R1-0528 have been released on HuggingFace, providing global developers with opportunities for in-depth research and optimization. In the future, with more test data being published and user feedback accumulating, R1-0528 is expected to demonstrate its potential in more fields.
AIbase believes that the release of DeepSeek R1-0528 is not only a technical breakthrough but also a bold attempt at AI popularization. Its powerful reasoning ability, optimized coding and writing functions, and free access model make this model one of the focal points of the AI field in 2025.
Open-source address: https://huggingface.co/deepseek-ai/DeepSeek-R1-0528