DeepSeek R1 0528 Qwen3 8B is an open-weights reasoning model released as part of the May 28, 2025, update to the DeepSeek-R1 series. It is a distilled version of the flagship DeepSeek-R1-0528 (a 685B parameter mixture-of-experts model), created by fine-tuning the Qwen3 8B base model on chain-of-thought (CoT) traces. This distillation process allows the smaller model to inherit complex reasoning patterns and logical depth typically found in much larger systems.
Capabilities and Performance
The model is optimized for high-intensity logical tasks, including mathematics, code generation, and multi-step reasoning. At its release, DeepSeek reported that the 8B variant achieved state-of-the-art results for its size class, notably outperforming the standard Qwen3 8B by approximately 10% on the AIME 2024 benchmark. It supports a context window of 128,000 tokens and is capable of generating up to 64,000 tokens in a single response.
In addition to reasoning, the model includes improved support for structured outputs such as JSON and functional calling. It also features architectural optimizations designed to reduce hallucinations during summarization and rewriting tasks compared to previous iterations of the R1 series. The model is released under the MIT License, permitting both academic research and commercial use.