DeepSeek R1 Distill Llama 70B

DeepSeek R1 Distill Llama 70B – Dense Reasoning Model via Knowledge Distillation

DeepSeek R1 Distill Llama 70B is a 70B dense model produced by fine-tuning Llama 3.3 70B Instruct on reasoning samples generated by DeepSeek R1. It is the largest of the six distilled models released alongside DeepSeek R1, and the only one in the set derived from the Llama 3.3 family — chosen officially for its stronger reasoning capability compared to Llama 3.1 at the same parameter scale.

Base Model & Licensing

Derived from Llama 3.3 70B Instruct, originally licensed under the Llama 3.3 license. Released under the MIT License by DeepSeek, which permits commercial use, modifications, and derivative works including further distillation for training other LLMs. The Llama 3.3 base license terms also apply given the model's derivation.

Usage Recommendations

Officially documented usage configurations for the R1 Distill series:

• Temperature: 0.5–0.7 (0.6 recommended) to prevent endless repetition or incoherent outputs
• System prompt: Avoid adding a system prompt; all instructions should be in the user prompt.