StorySoftware
DeepSeek updates R1 to R1-0528, raising its AIME 2025 score from 70.0 to 87.5
DeepSeek released R1-0528, an update to its open-weight reasoning model that raised AIME 2025 score from 70.0 to 87.5 and GPQA from 71.5 to 81.0. It reduced hallucinations and added JSON output and function calling; weights released on Hugging Face under MIT licence.
- AIME 2025: improved from 70% to 87.5%; GPQA: 71.5% to 81.0%.
- Performance on par with OpenAI o1; reduced hallucinations and enhanced reasoning.
- Added JSON output and function calling support.
- Open-weight under MIT licence on Hugging Face; no API changes for existing integrations.