DeepSeekReleased on Jan 20, 2025

DeepSeek R1 Zero: Benchmarks, Pricing & Size

DeepSeek R1 Zero is a language model from DeepSeek, released in January 2025.

DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. With RL, DeepSeek-R1-Zero naturally emerged with

DeepSeek R1 Zero model size

DeepSeek R1 Zero has 671 billion parameters and was trained on 14.8 trillion tokens. See how it compares to other models in the same parameter range.

ParametersTraining tokens
671B
14.8Ttokens
22× tokens-to-params ratio
Frontier (200B+)
671B
1B7B70B405B

DeepSeek R1 Zero API

API access coming soon

DeepSeek R1 Zero will be available through our gateway shortly.

DeepSeek R1 Zero examples

Recent arena outputs from DeepSeek R1 Zero, picked from the highest-ranked matchups.

DeepSeek R1 Zero license

DeepSeek R1 Zero is released under the MIT license, which permits commercial use, has 671.0B parameters.

License
MIT
Commercial use allowed
Parameters
671.0B

MIT License - allows commercial use

FAQ

Common questions about DeepSeek R1 Zero.

When was DeepSeek R1 Zero released?

DeepSeek R1 Zero was released on January 20, 2025 by DeepSeek. This is the official DeepSeek R1 Zero release date tracked on LLM Stats.

Is DeepSeek R1 Zero available via API?

Yes, DeepSeek R1 Zero is available via API. See the official documentation for authentication and endpoint details.

How big is DeepSeek R1 Zero?

DeepSeek R1 Zero has 671 billion parameters. It was trained on 14.8 trillion tokens. It ships as an open-weight model, so you can download and run it on your own hardware.

Who created DeepSeek R1 Zero?

DeepSeek R1 Zero was created by DeepSeek.

What is the license for DeepSeek R1 Zero?

DeepSeek R1 Zero is released under the MIT license. This is an open-source / open-weight license that permits self-hosting.

Where is the DeepSeek R1 Zero paper or technical report?

DeepSeek R1 Zero has a paper or technical report available at https://arxiv.org/abs/2501.12948. Use that source for architecture, training, release and evaluation details.