DeepSeek-R1 — open source GitHub project

DeepSeek-R1 is DeepSeek’s open reasoning-model series, including R1, R1-Zero, and distilled variants based on Qwen and Llama.

What it is

DeepSeek-R1 is the repository for DeepSeek’s reasoning models. It centers on DeepSeek-R1-Zero, DeepSeek-R1, and several dense distilled models based on Qwen and Llama.

R1-Zero demonstrates reinforcement learning without a preliminary supervised fine-tuning stage, while R1 adds cold-start data and extra tuning stages to improve readability and stability.

What is inside and how people use it

Inside are model descriptions, Hugging Face links, parameter tables, evaluation results, the paper, usage recommendations, and the list of distilled variants. The full R1 model is based on DeepSeek-V3-Base and uses a MoE architecture.

R1 family map

This fragment shows that the repository contains a family of base and distilled variants, not just one model.

Language: Markdown

DeepSeek-R1 family
- DeepSeek-R1-Zero: RL-first reasoning model
- DeepSeek-R1: cold-start data + RL/SFT stages
- Distill variants: Qwen and Llama based checkpoints
- Full models: 671B total, 37B activated, 128K context

In practice, R1 is often used through APIs, hosted inference services, or smaller distilled models. Running the full model locally requires serious infrastructure.

Strengths and limitations

The strength is openness around the reasoning pipeline and the availability of distilled variants. The repository shows not only a final model, but a research direction around RL and reasoning.

The limitation is size and evaluation complexity. A reasoning model can be strong at math and code, but a product still needs fact checking, safety work, and inference cost awareness.