← All models

DeepSeek-R1-Distill-Llama-70B

Published Publisher DeepSeek

Model overview

DeepSeek-R1-Distill-Llama-70B is a dense reasoning model with approximately 70B parameters released by DeepSeek. Based on Llama-3.3-70B-Instruct, it was fine-tuned using samples generated by DeepSeek-R1. It is a member of the R1 distilled model series, not the full DeepSeek-R1 model. (github.com)

Capabilities and uses

The model is intended for text generation and reasoning tasks and can be used for mathematical problem-solving, code-related tasks, and reasoning research. In DeepSeek's official evaluation report, this version achieved pass@1 scores of 70.0% on AIME 2024 and 94.5% on MATH-500, respectively; these results reflect only the corresponding evaluation conditions and do not guarantee performance on real-world tasks. (huggingface.co)

Limitations and deployment

The model's output may contain errors, repetition, or incoherence; important results should be checked by a human. Official guidance recommends setting temperature to 0.5–0.7 when running the R1 series, with 0.6 recommended; the model configuration and tokenizer should use the officially provided settings. Model weights are available for local deployment, and the resources required depend on precision, context length, and inference configuration. Official documentation states that the model weights are licensed under MIT, while noting that this version derives from a base model subject to the Llama 3.3 license; applicable license terms should be checked before use. (github.com)

Date

2025-01-20: Official listing and series release date. DeepSeek announced that day that it had simultaneously made six distilled models available, including the 70B model; the official model repository's commit history for this specific model also shows file uploads and release commits dated 2025-01-20. The date used here is the documented official listing date. (api-docs.deepseek.com)

Source note: In the current data, whether this specific model was first released before 2025-01-20 has not been independently verified; the date above is based on the official announcement and this model's repository commit history and is not presented as a verified earlier initial release date. (api-docs.deepseek.com)

Pricing

As of 2026-11-20, DeepSeek's official domestic pricing page in Chinese yuan (CNY) and international pricing page in US dollars (USD) did not list separate input and output token prices for this specific distilled model, so its official API pricing remains unverified. The DeepSeek-R1 API prices in the official 2025 announcement apply to the R1 service and cannot be used as a quote for this distilled model. Self-deployment also requires consideration of costs such as computing resources. (api-docs.deepseek.com)

Official sources