DeepSeek-R1-Distill-Llama-70B

DeepSeek-R1-Distill-Llama-70B is the flagship distilled model from the January 2025 DeepSeek-R1 release. It takes Meta's Llama-3.3-70B-Instruct as the base and fine-tunes it on 800,000 reasoning traces curated with DeepSeek-R1, producing a 70B model that reproduces much of R1's step-by-step reasoning behavior without the cost of the 671B mixture-of-experts parent.

Licensing is nuanced and flagged honestly here: DeepSeek publishes the repository under the MIT License, but the model is derived from Llama-3.3-70B-Instruct, so Meta's Llama 3.3 Community License terms also apply to the weights. It is aimed at teams that want frontier-class reasoning in a single-node 70B deployment, served with vLLM or SGLang with tensor parallelism.