DeepSeek releases R1 reasoning model and six smaller versions

DeepSeek has released R1, a reasoning model available through its chat website, API and downloadable weights. The 20 January 2025 launch also includes six smaller distilled models.

Users can select DeepThink on the website or call deepseek-reasoner through the API. Announced API prices are US$0.14 per million cached input tokens, US$0.55 for uncached input and US$2.19 for output.

DeepSeek-R1 at launch: chat website with DeepThink, API with deepseek-reasoner, or downloadable weights. — Three launch access routes. Graphic: Sulat.

Training and smaller models

The launch documentation describes R1 alongside R1-Zero, an experiment using reinforcement learning without preliminary supervised fine-tuning. DeepSeek says Zero developed reasoning behaviours but also suffered repetition, poor readability and language mixing. R1 adds initial training examples and combines supervised fine-tuning with reinforcement learning.

The main model has 671 billion parameters, with 37 billion activated. The six smaller models range from 1.5 billion to 70 billion parameters and use Qwen or Llama foundations, fine-tuned with R1-generated data. They offer a different scale of deployment from the main model.

Check the licence for each model

DeepSeek releases its main model weights and repository code under the MIT licence, permitting commercial use and modification. Its licensing notes also identify the distilled models’ underlying terms: Apache 2.0 for the Qwen-derived versions, Llama 3.1 for the 8B version and Llama 3.3 for the 70B version. Developers should check those base-model obligations before redistributing a derivative.

Featured illustration: Sulat. Official logo: DeepSeek.