DeepSeek-R1-Distill-Qwen-32B

DeepSeek-R1-Distill-Qwen-32B, released January 20, 2025 alongside the R1 family, is what happens when you fine-tune Qwen2.5-32B on reasoning traces from the 671B DeepSeek-R1. It was the strongest of the six official distills, beating OpenAI's o1-mini on math, coding, and science benchmarks despite being a fraction of the size — and it runs on a single high-end GPU.

Weights are MIT licensed, and as a dense transformer (not MoE) it is simpler to fine-tune and deploy than the flagship. It became one of the most replicated open reasoning models of 2025, mirrored across Hugging Face and served by Together AI, Groq, and others. Limitations: it inherits Qwen2.5-32B's 128K context ceiling in practice, and later dedicated reasoning models have since matched or passed its scores.