The newest update to the flagship reasoning model from Chinese AI startup DeepSeek, released on May 28, 2025. This update is making significant waves in the AI community, positioning DeepSeek as a formidable competitor to industry leaders like OpenAI and Google.It enhances performance in mathematics, programming, and logical reasoning, with notable improvements in benchmark tests. For instance, its accuracy on the AIME 2025 test increased from 70% to 87.5%, and its pass rate on LiveCodeBench improved from 63.5% to 73.3%. Additionally, the model’s performance on the “Humanity’s Last Exam” benchmark more than doubled, rising from 8.5% to 17.7% .

DeepSeek has distilled its large R1-0528 model into a much smaller version (Qwen3-8B), which matches or exceeds the performance of much larger models on certain benchmarks. This smaller model can run on a single GPU, making advanced AI more accessible to researchers and developers with limited hardware

Deepseek