DeepSeek-R1-Distill-Qwen-7B

DeepSeek-R1-Distill-Qwen-7B is a 7-billion-parameter reasoning model released by DeepSeek on January 20, 2025, alongside the R1 launch. It distills the reasoning patterns of the 671B DeepSeek-R1 into Qwen2.5-7B using 800K curated reasoning samples, producing a small model that emits an explicit <think> chain before its answer.

It was one of six distilled variants that brought R1-style test-time reasoning to consumer hardware, and it became a standard baseline for local reasoning agents. The weights carry the permissive MIT license, allowing commercial use, modification, and further distillation.