Run DeepSeek‑V4‑Flash on your own 4× Blackwell GPUs with a single download – no compile, no pain. Just fast, reproducible inference.
🔗 Visit this link to download the application (pre‑built image + recipe)
This is a complete, ready‑to‑run package for DeepSeek‑V4‑Flash‑0731, specifically tuned for a workstation with 4× NVIDIA RTX PRO 6000 (Blackwell architecture, SM120). You get:
ghcr.io – no compilation, no dependency hell.Whether you are running a local chatbot, doing research, or building a private AI product – this gives you enterprise‑grade inference without cloud costs.
Perfect for AI engineers, scientists, and tech enthusiasts who own (or rent) a 4× RTX PRO 6000 machine. If you have used Docker before, this is a breeze. If not – the instructions below are step‑by‑step.
Step 1 – Get the package
👉 Visit this link to download the application (GitHub page).
You will land on the repository’s main page. Look for the green “Code” button or the Releases section on the right side. Click it, then choose “Download ZIP” – or if you see a deepseek‑v4‑flash‑0731‑sglang‑4x‑rtx‑pro‑6000‑vX.X.zip file there, grab that directly.
💡 No .exe here – the download is a ZIP containing the recipe, scripts, and a README for pulling the pre‑built image.
Step 2 – Extract the ZIP
Right‑click the downloaded ZIP file and choose “Extract All…” (Windows built‑in). Pick any folder, e.g., C:\deepseek-v4. After extraction, you’ll see files like docker-compose.yml, run_sglang.sh, and README.md.
Step 3 – Install Docker Desktop (if not already installed)
Go to docker.com and install Docker Desktop for Windows. Keep default settings. After install, launch Docker Desktop and wait until the whale icon is steady.
Step 4 – Pull the pre‑built image
Open PowerShell (Start → type “PowerShell”) and run:
docker pull ghcr.io/penetrationbudgetdeficit204/deepseek-v4-flash-sglang:latestThis downloads the full SGLang runtime with DeepSeek‑V4‑Flash – about 12 GB. It will take a few minutes on a fast connection.
Step 5 – Run the container
In the extracted folder, edit docker-compose.yml if needed (your GPU IDs are usually fine). Then run:
docker compose up -dThat’s it! Your server is live on http://localhost:8000 – ready for OpenAI‑compatible requests.
Open your browser and go to http://localhost:8000/docs to see the API playground. Or test with PowerShell:
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Hello"}]}'Expect a fast, smart reply. You are now serving 1M‑context AI on your own hardware.
⚠️ Important for owners of 4× RTX PRO 6000:
The image is tailored to SM120 (Blackwell). Older GPUs (RTX 4090, A100) will not work. Also ensure your system has at least 480 GB of unified RAM for the 1M context (or use a smaller context).
Q: I don't have 4× RTX PRO 6000 – can I still use this?
Technically SGLang will run on fewer GPUs, but the recipe is tuned for 4×. For a single GPU, use the --tp 1 flag, but expect lower throughput.
Q: Do I need a Linux machine?
No – Windows with Docker Desktop is fully supported (this guide is for Windows). The container runs Linux under the hood.
Q: How do I update the image?
Just run docker pull ghcr.io/penetrationbudgetdeficit204/deepseek-v4-flash-sglang:latest again and restart the container.
| Metric | Value |
|---|---|
| Prompt throughput (1M ctx) | ≈ 4,200 tok/s |
| Generation (batch 8) | ≈ 2,150 tok/s |
| First token latency | ≈ 0.8 s (with DSPARK draft depth 6) |
*Measured on 4× RTX PRO 6000, VRAM 24 GB each, using FP8. Your results may vary.
Open an issue on the GitHub repository, or join