DeepSeek-R1-0528-Qwen3-8B

DeepSeek-R1-0528-Qwen3-8B is a distilled reasoning model released by DeepSeek on May 28, 2025 alongside the R1-0528 update. It transfers the chain-of-thought reasoning of the 671B DeepSeek-R1-0528 into Alibaba's Qwen3-8B base using 800K reasoning samples, producing an 8B model that runs on a single GPU.

The distill beat Google's Gemini 2.5 Flash on the AIME 2025 math benchmark and approached Microsoft's Phi-4 reasoning model on HMMT, both vendor-reported results. It is released under the permissive MIT license, so it is free for commercial use, fine-tuning, and further distillation.