DeepSeek has just announced a powerful upgrade to its flagship AI model, DeepSeek R1, unveiling version DeepSeek-R1-0528. This new release represents a leap forward in reasoning depth, coding ability, and math-solving skills, pushing the model closer than ever to matching giants like OpenAI's O3 and Gemini 2.5 Pro.
Thanks to enhanced post-training algorithms and more computing power, DeepSeek-R1-0528 has made major progress in handling complex reasoning tasks. For example, in the notoriously tough AIME 2025 math test, the model's accuracy jumped from 70% to 87.5%. It now uses nearly double the thinking power per question, averaging 23,000 tokens versus 12,000 previously, a sign of deeper and more nuanced reasoning.
The model shines across a wide range of benchmarks:
Beyond performance, users will enjoy a lower hallucination rate, smoother function calling, and a more intuitive experience when using the model for coding support (also known as "vibe coding").
Interactive comparison between original and improved versions
The update also brings something for open-source fans: a distilled version called DeepSeek-R1-0528-Qwen3-8B. It transfers the thinking patterns of DeepSeek-R1 into the lightweight Qwen3 8B model, achieving state-of-the-art resultsamong models in its class. It even rivals much larger models like Qwen3-235B on tests like AIME 2024.
You can interact with DeepSeek-R1-0528 at chat.deepseek.com by activating the "DeepThink" mode, or access it via API at platform.deepseek.com. For developers, running it locally is also easier than ever. The latest version now supports system prompts and no longer requires special formatting to trigger advanced reasoning.
All DeepSeek-R1 models, including this latest version, are released under the MIT License, allowing for full commercial use and distillation. This means startups, researchers, and developers can build on top of the model freely.