← Back to Discover

DeepSeek-R1

Open-source reasoning model family from DeepSeek: 671B-parameter mixture-of-experts checkpoints trained with reinforcement learning, plus six smaller distilled models, usable through web chat, an API or your own servers.

Open modelsModel APIs
Visit website ↗

What it does

DeepSeek-R1 is the first generation of DeepSeek's reasoning models, published as open weights under the MIT License. The repository covers two main models, DeepSeek-R1-Zero and DeepSeek-R1. R1-Zero was trained with reinforcement learning alone, with no supervised fine-tuning step beforehand. R1 adds cold-start data ahead of the reinforcement learning to deal with problems the page lists for R1-Zero: repetition, hard-to-read output and mixed languages. Both flagship models are mixture-of-experts models with 671B total parameters, 37B activated parameters and a 128K context length. DeepSeek also publishes six dense models fine-tuned on reasoning data generated by R1, at 1.5B, 7B, 8B, 14B, 32B and 70B parameters, built on Qwen2.5 and Llama 3 base models. The repository reports benchmark results for math, code and general reasoning against models such as OpenAI o1 and GPT-4o. You can chat with the model on DeepSeek's official website by switching on DeepThink, or call it through an OpenAI-compatible API on the DeepSeek Platform. For local use, the distilled models run like Qwen or Llama models, and the page gives example vLLM and SGLang commands. Hugging Face Transformers does not support the full R1 models directly yet, and the page points to the DeepSeek-V3 repository for running them. Recommended settings include a temperature around 0.6 and keeping all instructions in the user prompt instead of a system prompt. The MIT License covers the code and weights, allows commercial use and permits derivative works, including distillation to train other models. The distilled models built on Qwen and Llama also carry the licenses of their base models.

Capabilities

  • 671B-parameter mixture-of-experts reasoning model with a 128K context length
  • R1-Zero trained with reinforcement learning only; R1 adds cold-start data
  • Six distilled dense models from 1.5B to 70B parameters
  • MIT-licensed code and weights, with commercial use allowed
  • Web chat with DeepThink and an OpenAI-compatible API

When to use it

  • You want an openly licensed reasoning model for math, code or multi-step problems
  • You need smaller distilled checkpoints you can serve yourself with vLLM or SGLang
  • You plan to distill or fine-tune your own models using R1 outputs
  • You want to try the model in the web chat before using the API