DeepSeek has released DeepSeek-R1-0528, a minor version upgrade that significantly enhances depth of reasoning and inference capabilities through increased computational resources and algorithmic optimizations during post-training.

  • AIME 2025 test accuracy increased from 70% to 87.5%, driven by an average of 23K tokens per question compared to the previous 12K.
  • Overall performance approaches leading models like O3 and Gemini 2.5 Pro in mathematics, programming, and general logic.
  • The update reduces hallucination rates and improves function calling support and vibe coding experiences.
  • DeepSeek distilled chain-of-thought from this model to post-train Qwen3 8B Base, creating DeepSeek-R1-0528-Qwen3-8B which achieves state-of-the-art performance among open-source models on AIME 2024.

The upgraded model offers better handling of complex reasoning tasks and is available via Hugging Face, local apps like vLLM and SGLang, and DeepSeek's official API.