DeepSeek-R1
Open-weight reasoning language models from DeepSeek, released for research and local use.
- Stars
- 92k
- Forks
- 11.7k
- Licence
- MIT
- Language
- —
- Last update
- 1.4 years ago
- Type
- AI model / agent
No commits in over a year — check for forks or alternatives before adopting.
What it is
DeepSeek-R1 is a family of large language models trained with reinforcement learning to work through problems step by step, aimed at math, code, and reasoning tasks. The README says the weights are openly released, along with six smaller dense models distilled from it based on Qwen and Llama. People use it to run or study a reasoning model without relying on a closed commercial API.
- Download the DeepSeek-R1 or DeepSeek-R1-Distill weights from Hugging Face
- Run the distilled checkpoints locally to test reasoning on math or coding prompts
- Use the reasoning data pattern as a reference for distilling smaller models
- Compare results against the benchmark table in the README
Great for
- Researchers studying how reinforcement learning produces reasoning behavior 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
Not ideal for
- Users with limited hardware, since the full 671B-parameter model is very large to run locally (the distilled 1.5B to 70B versions are smaller options)
- People who need a polished chat app with no setup; this repository is model release material, not a finished product
- Anyone expecting the repository itself to contain runnable code, since the README covers model downloads and evaluation rather than a software package
How to get it
- 1
See the README
You'll need: Hugging Face account or access to the model weights · Sufficient GPU memory for the chosen model size
Clone the repository
git clone https://github.com/deepseek-ai/DeepSeek-R1.git - 2
Run it
- Read the Usage Recommendations section of the README before running any model
- Pick a checkpoint from the model download tables that fits your hardware
- Download the weights from the linked Hugging Face page
- Load the model with a compatible inference tool and test a reasoning prompt
- 3
Get the most out of it
- Follow the README's Usage Recommendations for settings, since the distilled models use modified configs and tokenizers
- Use the evaluation settings noted in the README (temperature 0.6, top-p 0.95, max 32,768 tokens) as a starting point for comparisons
- Start with a distilled 1.5B or 7B checkpoint to test before moving to larger models
- Read the linked DeepSeek_R1.pdf paper to understand the training pipeline
Downloads
Strengths
- Openly released weights under the MIT license, with the paper linked from the README
- The README reports benchmark results against models such as OpenAI o1 and Claude-3.5-Sonnet
- Multiple sizes are available, from 1.5B to 70B distilled checkpoints
Trade-offs
- The full model has 671B total parameters, which requires substantial hardware to run
- The README itself notes that R1-Zero had issues with endless repetition, poor readability, and language mixing, which motivated the cold-start approach in R1
- Benchmark numbers are self-reported by the project and not independently verified here
Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.