# deepseek-ai/deepseek-r1 architecture

> Architecture diagram of the GitHub repository deepseek-ai/deepseek-r1, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-11-04). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/deepseek-ai/deepseek-r1
- Repository: https://github.com/deepseek-ai/deepseek-r1

## Overview

This repository is a research and model-release landing page for DeepSeek-R1, not an implementation or serving runtime: its tree contains a README, the paper PDF, and figures, with no source-code modules. The documented lifecycle spans base-model RL, cold-start/SFT and RL post-training, reasoning-data distillation into dense models, and evaluation. Links among these stages are stated in the README; no implementation-level calls or data paths can be established from the available files.

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_training["Model development"]
  node_base["V3 base model"]
  node_zero["R1-Zero RL"]
  node_coldstart["Cold-start data"]
  node_r1pipeline["R1 SFT and RL"]
  node_reasoningdata["Reasoning samples"]
  node_densefine["Dense-model fine-tuning"]
end

subgraph group_release["Models and evaluation"]
  node_r1models["R1 models"]
  node_distilled["Distilled checkpoints"]
  node_evaluation["Benchmark evaluation"]
end

node_researcher(("Researcher"))
node_community(("Research community"))

node_researcher -.->|"starts from"| node_base
node_base -->|"trains with RL"| node_zero
node_coldstart -->|"seeds training"| node_r1pipeline
node_base -->|"trains with SFT and RL"| node_r1pipeline
node_r1pipeline -->|"generates samples"| node_reasoningdata
node_reasoningdata -->|"fine-tunes"| node_densefine
node_densefine -->|"produces"| node_distilled
node_r1pipeline -.->|"releases"| node_r1models
node_r1models -->|"evaluated on"| node_evaluation
node_distilled -->|"evaluated on"| node_evaluation
node_community -->|"accesses"| node_r1models
node_community -->|"downloads"| node_distilled

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_base,node_zero,node_coldstart,node_r1pipeline,node_reasoningdata,node_densefine toneBlue
class node_r1models,node_distilled,node_evaluation toneAmber
class node_researcher,node_community toneIndigo
```

## Components

### Model development

- **V3 base model**
- **R1-Zero RL**
- **Cold-start data**
- **R1 SFT and RL**
- **Reasoning samples**
- **Dense-model fine-tuning**

### Models and evaluation

- **R1 models**
- **Distilled checkpoints**
- **Benchmark evaluation**

### Other components

- **Researcher**
- **Research community**

## Connections

- Researcher → V3 base model: starts from (evidence: `README.md`)
- V3 base model → R1-Zero RL: trains with RL (evidence: `README.md`)
- Cold-start data → R1 SFT and RL: seeds training (evidence: `README.md`)
- V3 base model → R1 SFT and RL: trains with SFT and RL (evidence: `README.md`)
- R1 SFT and RL → Reasoning samples: generates samples (evidence: `README.md`)
- Reasoning samples → Dense-model fine-tuning: fine-tunes (evidence: `README.md`)
- Dense-model fine-tuning → Distilled checkpoints: produces (evidence: `README.md`)
- R1 SFT and RL → R1 models: releases (evidence: `README.md`)
- R1 models → Benchmark evaluation: evaluated on (evidence: `README.md`)
- Distilled checkpoints → Benchmark evaluation: evaluated on (evidence: `README.md`)
- Research community → R1 models: accesses (evidence: `README.md`)
- Research community → Distilled checkpoints: downloads (evidence: `README.md`)
