Deepseek, the Ai Unicorn Chinese, published an updated version of its R1 reasoning model, named Deepseek-R1-0528. This version improves the capabilities of the model in mathematics, programming and general logical reasoning, positioning it as a formidable open source alternative to the main models like O3 of Openai and Gemini 2.5 Pro from Google.
Technical improvements
Update R1-0528 introduces significant improvements in the depth of reasoning and inference accuracy. In particular, the performance of the model on the mathematical reference likes 2025 increased from 70% to 87.5%, reflecting a deeper reasoning process which has an average of 23,000 tokens per question, against 12,000 in the previous version. This improvement is attributed to an increase in calculation resources and algorithmic optimizations applied during post-training.
In addition to mathematical reasoning, the model has shown improved performance in code generation tasks. According to LiveCodebench Benchmarks, R1-0528 ranks just below the O4 Mini and O3 models of Openai, outperforming the Grok 3 Mini 3 of XAI and Qwen 3 of Alibaba in code generation tasks.
Open Source Model Weight
Deepseek continues its commitment in terms of opening and open weights AI by releasing R1-0528 under the MIT license, allowing developers to freely modify and deploy the model. The weights of the model are available on the face of hugs, and detailed documentation is provided for local deployment and the integration of the API. This approach contrasts with the nature owner of many leading AI models, promoting transparency and accessibility in the development of AI.
Distilled model for light deployment
Recognizing the need for more accessible AI solutions, Deepseek also published a distilled version of R1-0528, named Deepseek-R1-0528-QWEN3-8B. This model, refined from the QWEN3-8B ​​of Alibaba using the text generated by R1-0528, achieves advanced performance among open source models on the reference likes 2025. It is designed to operate effectively on a single GPU, which makes the advanced capacities of AI more accessible to developers with limited calculation resources.
Censorship considerations
While Deepseek's progress in AI is remarkable, the R1-0528 model has been observed as a stricter content moderation compared to its predecessors. Independent tests have revealed that the model avoids or provides limited responses to politically sensitive subjects, such as the manifestations of Tiananmen Square and the status of Taiwan, aligning on Chinese regulations which oblige AI models to respect content restrictions.
Global implications
The publication of R1-0528 underlines the growing influence of China in the AI ​​sector, which questions the domination of American companies. Deepseek's ability to develop competitive AI models to a fraction of the cost of their Western counterparts caused corporate responses like Openai, who expressed concerns about the potential for manipulating these models by the Chinese government. This development highlights the changing dynamics of global development of AI and the growing importance of open source models in promoting innovation and competition.
Conclusion
Deepseek's R1-0528 model represents a significant increase in open source AI, offering improved reasoning capacities and accessibility for developers. By providing both a large-scale model and a distilled version adapted to the single-GPU deployment, Deepseek makes progress in the democratization of AI technology. However, the membership of the model to content moderation policies reflects the complex interaction between technological advancement and regulatory compliance. While the AI ​​landscape continues to evolve, Deepseek's developments will probably play a central role in training the future of Open Source.
Discover the Open Source And Try it now. All the merit of this research goes to researchers in this project. Also, don't hesitate to follow us Twitter And don't forget to join our 95K + ML Subdreddit and subscribe to Our newsletter.
Asif Razzaq is the CEO of Marktechpost Media Inc .. as a visionary entrepreneur and engineer, AIF undertakes to exploit the potential of artificial intelligence for social good. His most recent company is the launch of an artificial intelligence media platform, Marktechpost, which stands out from its in-depth coverage of automatic learning and in-depth learning news which are both technically solid and easily understandable by a large audience. The platform has more than 2 million monthly views, illustrating its popularity with the public.
