← All models

DeepSeek-R1-Zero

Published Publisher DeepSeek

Model Overview

DeepSeek-R1-Zero is a reasoning large language model from DeepSeek, based on DeepSeek-V3-Base and trained through large-scale reinforcement learning without prior supervised fine-tuning. It is a different version from the subsequent DeepSeek-R1. (huggingface.co)

Capabilities and Uses

The model is designed for complex reasoning and can be used for mathematical problems, code, and research into reasoning methods. The official model card lists 671B total parameters, 37B activated parameters per activation, and a context length of 128K; the model weights are available under the MIT license and can be downloaded and deployed. (huggingface.co)

Limitations

The official source notes that DeepSeek-R1-Zero may exhibit repeated repetition, poor output readability, and language mixing. Generated results still need to be verified in the context of the specific task. (huggingface.co)

Date

Record date: 2025-01-20, based on the listing and file commit dates of the official model repository for this version. The repository commit history shows that an initial commit and model file uploads had already occurred on that date. (huggingface.co)

Source note: In the current data, the standalone initial release date of DeepSeek-R1-Zero has not been independently verified. The date above is the listing date recorded in the model repository and is not an independently confirmed initial release date. Official commit history (huggingface.co)

Pricing

Lookup date: 2026-11-20. DeepSeek's official Chinese pricing page and English pricing page were searched, but neither contained API input or output prices explicitly applicable to DeepSeek-R1-Zero, so pricing has not been verified. DeepSeek-R1 pricing cannot be treated as R1-Zero pricing. (api-docs.deepseek.com)

Official Sources