DeepSeek-R1-Distill-Qwen-32B
A Qwen-based distillation of DeepSeek-R1's reasoning traces into a smaller open model
DeepSeek-R1-Distill-Qwen-32B is a dense 32-billion-parameter model that DeepSeek built by fine-tuning Qwen2.5-32B on reasoning traces generated by the full DeepSeek-R1 model. Rather than training a smaller model from scratch, DeepSeek distilled R1’s chain-of-thought outputs into several base models, and the 32B Qwen version turned out to be the strongest of the group. It beat OpenAI’s o1-mini on math, coding, and science benchmarks despite its comparatively modest size, which made it one of the first open models small enough to run on a single high-end GPU while still showing real reasoning ability. The weights are released under the MIT license alongside the rest of the R1 family, and the model is widely mirrored on Hugging Face and served by inference providers such as Together AI and Groq.
It fits a specific niche: teams that want R1-style step-by-step reasoning without paying for or hosting the 671B full model. Because it’s a standard dense transformer rather than a mixture-of-experts model, it’s also simpler to fine-tune and deploy than DeepSeek’s flagship releases.