Run Qwen
on your own machine

Swap api.openai.com for localhost.
One command. Your hardware, your data, no bill per token.

bash

$ curl -fsSL https://github.com/Ga0512/SursumAI/raw/v1.0.18/install.sh | bash

1command to install

0bytes of your prompts leaving

4Python packages, all pinned

MITthe whole app, on GitHub

Drop-in

Your code does not change

It speaks the OpenAI API. Point it at your own machine and carry on.

Before

client = OpenAI(
    base_url="https://api.openai.com/v1",
    api_key="sk-proj-...",
)

client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[...],
)

After

client = OpenAI(
    base_url="http://localhost:8001/v1",
    api_key="sk-sursum-...",
)

client.chat.completions.create(
    model="Qwen/Qwen3-8B-GGUF",
    messages=[...],
)

Streaming, JSON schema, vision, tools — the same fields you already send. See the API docs →

Nothing you send leaves your machines. Swap the GPU, add another one, and the URL your code calls stays the same.

Deploy without knowing any of it

Pick a model, press Deploy. The GPU, the runtime, the port and the keys are decided for you — the decisions you would otherwise take alone, at 2am.

Quick start →

The dashboard listing deployments with their status

Watch it like a service, not a script

Tokens per second, memory, logs. When something dies, the card says what happened in one sentence and what to press.

When something breaks →

Try it before your code does

A playground on the real endpoint, images included. The reasoning folds away; the answer does not.

Reasoning models →

The playground chatting with a deployed model

Copy the snippet, not the concept

Python, JavaScript and curl, already filled in with your URL and your key.

API docs →

Ready-made code snippets for the deployed model

Too big for this machine?

Paste the SSH line of any server and it installs itself there, behind a private tunnel. Same dashboard, same key, same URL — your code never learns the model moved.

How machines work →

A pool mixing models from different machines

Measured

Numbers, and where they came from

No vendor slides. Two we kept coming back to:

1.9s to pull six fields out of a 25 000-character document, four running at once — Qwen3.8 27B on an RTX 6000 Ada

5.9× less output with reasoning off on the deployment: 1175 tokens became 176, same answer

Measured while building this, on rented GPUs. The cards, the settings and the arithmetic that predicts them are in the docs — and the router benchmark, 100 GSM8K problems with every answer checked, is in the repository.

Why

The hard way vs. this

Renting an API

Pay per token, forever
Your prompts leave the building
Prices and models change under you

Doing it yourself

A day of Docker, CUDA and VRAM maths
Out of memory at 2am
One model, one machine, no API

SursumAI

One command
An OpenAI URL and a key
Nothing leaves your machine
Too big for it? Rent a GPU — same URL

Ollama runs one model. LiteLLM routes you to someone else’s. This is the provider, and it is yours.

Pricing

Everything on your own machine is free, forever. Pro is a subscription for running models on the servers you already have.

Free

$0forever

Unlimited models on your machine
Router, pools and playground
OpenAI-compatible API with account keys
GPU detected and used automatically
Open source (MIT)
Community support on GitHub

Install free

Pro Beta

$15/ month — or $120 / year

Everything in Free
Deploy to your own servers over SSH — unlimited machines
Same install, unlocked by a token — nothing to reinstall
One dashboard and one API for the models on all of them
Nothing exposed on those machines: a private SSH tunnel
Email support
Cancel anytime — Pro keeps working until the end of what you paid

Go Pro

No per-user or per-token pricing: your models run on your hardware, and your prompts never pass through us.

FAQ

Questions people ask first

Does my data really stay here?: Yes. The models run on hardware you control and the prompts never reach us. The only thing our servers know is whether an account subscribes — the free edition does not talk to us at all.
What do I need to run it?: Any Linux, macOS or Windows with WSL. An NVIDIA GPU makes it fast; without one it still runs on the CPU, slower. The installer brings everything else.
Is it really open source?: MIT, the whole app. Pro is a subscription that unlocks running models on your other servers over SSH; everything else is free forever and always will be.
What happens if you disappear?: Nothing stops. The app is on your machine and the code is on GitHub. Even Pro keeps working for 30 days offline, because the answer about your account is signed and dated.

Get started

Install in one command

Runs on Linux, macOS and WSL (Windows — open Ubuntu, not PowerShell). No Docker, Python or GPU knowledge needed. The same command installs Free and Pro.