DS4 with Docker: Step-by-Step Installation Tutorial for DeepSeek 4

Step-by-Step: Run DeepSeek 4 Locally with Docker

This tutorial walks you through a complete DS4 deployment with Docker. In about ten minutes you will have the DeepSeek 4 Flash model answering prompts from a local, OpenAI-compatible API — completely offline.

Step 1 — Prerequisites

• Docker Engine 20.10+ (or Docker Desktop)
• At least 4 GB RAM (8 GB recommended) and a GPU with Metal, CUDA or ROCm support
• ~10 GB free disk space for the model weights

Step 2 — Start the Container

Create a working directory and a docker-compose.yml:

services:
  ds4:
    image: arraying/ds4:latest
    restart: unless-stopped
    ports:
      - 8080:8080
    volumes:
      - ./data/ds4:/data

Then start it: docker compose up -d. On first launch the container downloads the DeepSeek 4 Flash weights automatically — grab a coffee while it pulls.

Step 3 — Send Your First Prompt

Once the container is up, test it with a plain HTTP request:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"ds4-flash","messages":[{"role":"user","content":"Explain how DS4 works in one sentence"}]}'

Because the API is OpenAI-compatible, you can point existing clients — LangChain, Open WebUI, custom scripts — at http://localhost:8080/v1 with no code changes.

Step 4 — Pick Your Model Size

DS4 runs both the Flash model (fast, lightweight — ideal for everyday workloads) and the PRO model (heavier, higher quality — for complex reasoning). Switch by changing the model field in your requests. For a preview of expected throughput on high-end Apple Silicon, antirez's benchmark charts are worth a look:

Common Pitfalls

That's it — you now run DeepSeek 4 Flash and PRO entirely on your own hardware. For the full deployment reference, including system requirements and prebuilt compose files, head to the DS4 tool page.