Guide · Open source
How to use DeepSeek-R1: download the models and run them
Updated · written with AI from the project's own documentation and checked against our catalogue data
DeepSeek-R1 is a family of open-weight reasoning models trained with reinforcement learning to work through math, code and reasoning problems step by step. This guide is for people who want to run or study a reasoning model themselves instead of calling a closed commercial API. By the end you will know what the repository contains, which model sizes exist, what you need before downloading, and where to look for evaluation details.
DeepSeek-R1deepseek-ai/DeepSeek-R1 · ★ 92k · MIT Install options & downloads Quick start
- 1Read the README. Start with the Usage Recommendation section and the model tables to choose a model.
- 2Check your GPU memory. Match a checkpoint size from 1.5B to 70B to the memory you have. The full 671B model is very large to run locally.
- 3Get Hugging Face access. Create a Hugging Face account or confirm you have access to the model weights.
- 4Clone the repository. Run git clone https://github.com/deepseek-ai/DeepSeek-R1.git to get the repository locally.
- 5Download your chosen checkpoint. Open the matching Hugging Face page from the README table and follow its download instructions.
- 6Apply the recommended settings. Use the configs and tokenizers from the model page and generation settings close to the README's evaluation setup.
What is in the DeepSeek-R1 repository?
The repository is model release material. The README describes the DeepSeek-R1-Zero and DeepSeek-R1 models, six dense models distilled from DeepSeek-R1, model download links, and evaluation results. It does not ship a finished chat application or a software package you install and run as a program.
The README says the model weights are openly released under the MIT licence.
- DeepSeek-R1-Zero and DeepSeek-R1: 671B total parameters, 37B activated, 128K context length, both linked to Hugging Face.
- Distilled checkpoints: 1.5B, 7B, 8B, 14B, 32B and 70B, based on Qwen2.5 and Llama3 series models.
Check your hardware and access before you download
The full 671B-parameter model is very large to run locally, so most readers will start with a distilled checkpoint. The README lists the smaller distilled options from 1.5B up to 70B. Choose a size that fits the GPU memory you have.
You need a Hugging Face account or access to the model weights, since the model files are linked through Hugging Face. The README also asks you to review its Usage Recommendation section before running the DeepSeek-R1 series locally.
- Sufficient GPU memory for the size you pick (the README does not give exact memory figures, so check the model card on Hugging Face).
- A Hugging Face account or access to the weights.
- Read the Usage Recommendation section of the README first.
Clone the repository
To get a local copy of the README and repository files, clone it with git using the command from the README.
git clone https://github.com/deepseek-ai/DeepSeek-R1.gitDownload a model from Hugging Face
Each model in the README tables links to its page on Hugging Face. Open the page for the checkpoint you chose and follow the download instructions there. For the distilled models, the README notes that they are fine-tuned from open-source base models using samples generated by DeepSeek-R1, and that the configs and tokenizers were slightly changed, so use the settings the authors provide.
Distilled models and their base models include DeepSeek-R1-Distill-Qwen-1.5B (based on Qwen2.5-Math-1.5B), DeepSeek-R1-Distill-Qwen-32B (based on Qwen2.5-32B) and DeepSeek-R1-Distill-Llama-70B (based on Llama-3.3-70B-Instruct).
- Pick the smallest checkpoint your hardware can handle, then move up.
- Use the configs and tokenizers from the model page, as the README asks.
Use the recommended generation settings
The README reports its evaluation settings: a maximum generation length of 32,768 tokens, and for sampled benchmarks a temperature of 0.6 and top-p of 0.95, with 64 responses per query to estimate pass@1. Using settings close to these makes your results easier to compare with the published numbers.
- Maximum generation length: 32,768 tokens (README evaluation setting).
- Temperature: 0.6 and top-p: 0.95 for sampling benchmarks.
Compare results and study the method
The evaluation table in the README compares DeepSeek R1 with DeepSeek V3, OpenAI o1-mini, OpenAI o1-1217, Claude-3.5-Sonnet-1022 and GPT-4o 0513 across benchmarks such as MMLU, GPQA-Diamond, LiveCodeBench and AIME 2024. The README also links the paper, DeepSeek_R1.pdf, which describes the training pipeline in more detail.
- Read the Model Summary section to see the two RL stages and two SFT stages.
- Read the paper linked from the README for the full method.
What people use it for
- Researchers studying how reinforcement learning produces reasoning behaviour in language models.
- Developers who want an open-weight reasoning model to run or fine-tune themselves.
- Teams with enough GPU memory to run the smaller distilled checkpoints locally.
- Students comparing reasoning models on math and coding benchmarks.
Frequently asked questions
Is DeepSeek-R1 free to use?
The repository is released under the MIT licence, and the README says the model weights are openly released. Check the LICENSE file in the repository for the exact terms that apply to you.
What hardware do I need to run DeepSeek-R1?
The full 671B-parameter model is very large to run locally. The README lists smaller distilled checkpoints from 1.5B to 70B, and you need enough GPU memory for the size you choose. The README does not give exact memory figures.
Do I need a Hugging Face account?
The README lists Hugging Face as a prerequisite, either an account or access to the model weights. The model download links point to Hugging Face pages.
Does the repository contain runnable code?
No. The README covers model downloads and evaluation rather than a software package, so the repository alone will not run a chat app for you.
What are the alternatives to DeepSeek-R1?
The README's comparison includes OpenAI o1 (closed, for comparison). Other projects people compare with it include Qwen QwQ and Llama models from Meta.
How does DeepSeek-R1 compare with OpenAI o1?
The README states that DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code and reasoning tasks. Its evaluation table gives benchmark-by-benchmark numbers you can check directly.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
DeepSeek-R1 suits researchers, developers and students who want open weights of a reasoning model and have the GPU memory for a distilled checkpoint. If you need a polished chat app with no setup, or you have limited hardware, pick a hosted product or a smaller model instead.
Install options & downloads