deepseek-ai /
DeepSeek-R1
A 671B open reasoning model that thinks step by step, plus smaller distilled versions.
- 14.4k likes
- 1.5M downloads a month
- 684B parameters
- Some experience
- Text & code
DeepSeek-R1 is a large language model with 671B total parameters, of which 37B are activated at a time, and a 128K context length. It is built on DeepSeek-V3-Base. Its predecessor, DeepSeek-R1-Zero, was trained with reinforcement learning (RL) and no supervised fine-tuning first; it reasoned well but had problems like endless repetition, poor readability and language mixing. DeepSeek-R1 adds cold-start data before RL to fix those issues, and the card says it performs comparably to OpenAI-o1 on math, code and reasoning tasks.
The card also releases six smaller dense models distilled from R1 (1.5B, 7B, 8B, 14B, 32B and 70B), based on Qwen2.5 and Llama3. Honest limits: the full model is very large, and the card says Hugging Face Transformers does not directly support it yet, so you are pointed to the DeepSeek-V3 repo for local running. It is not best at everything in the card's table; for example, it scores lower than some others on IF-Eval and SimpleQA. The card also warns that the model can skip its thinking step, and that it needs particular settings to behave well.
What you need to run it
684B parameters. The weights take about 1369 GB at full 16-bit precision (fits a multi-GPU server), or about 411 GB as a 4-bit download, which runs on a multi-GPU server. Long inputs and images need more memory on top.
What it does well
Strong math and code results
In the card's tests R1 scores 79.8 on AIME 2024 and 97.3 on MATH-500, with a Codeforces rating of 2029 and 65.9 on LiveCodeBench.
Smaller distilled models
Six distilled checkpoints (1.5B to 70B) are released. The card says DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks.
Reasoning learned through RL
The card says R1-Zero is the first open research to show reasoning can be incentivized purely through reinforcement learning, without supervised fine-tuning first.
Several ways to try it
You can chat on chat.deepseek.com with the DeepThink button, use an OpenAI-compatible API at platform.deepseek.com, or download the weights.
Permissive licence
MIT licence, covering commercial use, modification and derivative works, including distillation for training other LLMs.
What you could build
- Work through maths competition-style problems and see the model's step-by-step reasoning.
- Get help with coding problems, using the reasoning to check logic before writing code.
- Run a distilled model (for example the 7B or 14B) on your own hardware to experiment with reasoning models.
- Use the OpenAI-compatible API to add a reasoning model to a small app.
- Study how reinforcement learning and distillation produce reasoning behaviour, using the model and its paper.
Get started
The easiest start is chat.deepseek.com with DeepThink switched on. To run a distilled model yourself, serve it with vLLM or SGLang using the commands below, and follow the usage recommendations (temperature around 0.6, no system prompt).
vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eagerpython3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --trust-remote-code --tp 2Commands are copied from the model card. Read it on Hugging Face for the full setup.
