DeepSeek-R1-Distill-Llama-8B

DeepSeek-R1-Distill-Llama-8B is DeepSeek's January 2025 distillation of R1's reasoning ability into Meta's Llama-3.1-8B-Base, fine-tuned on 800K reasoning samples curated from DeepSeek-R1. It achieves 89.1% pass@1 on MATH-500, far beyond typical 8B models.

Released under the MIT license, it made chain-of-thought reasoning practical on a single consumer GPU and spawned a wave of small reasoning models and local agent stacks.