Copy-paste three commands on your computer or server. No tech skills needed — just follow along.
Sovnex runs on Docker. Set up what your system needs below and keep them running. Already have them? Skip this.
1. Docker Desktop — download, install and open it (the whale icon in the menu bar means it's ready).
docker.com/products/docker-desktop ↗2. Ollama — the local AI engine. Download the app, install and open it (it sits in the menu bar); Sovnex connects to it automatically. Skipped it? The installer offers to install it with Homebrew and start it — without Homebrew you have to install it yourself.
ollama.com/download ↗Paste this to auto-install Docker (needs sudo):
$ curl -fsSL https://get.docker.com | sh $ sudo systemctl enable --now docker
Ollama needs nothing here — the installer sets the local AI engine up inside Docker (uses your NVIDIA GPU if present, otherwise the CPU). On very large NVIDIA servers it runs vLLM instead and creates no Ollama container.
Linux shortcut: you can skip Docker here and go to step 2 — ./install.sh will offer to install Docker for you (also needs sudo).
Sovnex picks its AI engine from your hardware — you do not choose.
| Your machine | What runs |
|---|---|
| NVIDIA GPU with 40GB+ VRAM, or a GB10 / DGX Spark class box | vLLM + a 27B model — the best answers this product gives |
| Anything else — laptop, CPU-only server, smaller GPU | Ollama + a 7B model — 16GB of RAM is enough |
Want to check before installing? Download the installer (step 1), then run ./install.sh --dry-run. It reads your machine and prints which engine it would use, which model, and how it would size everything — without changing anything on the computer.
The installer is available to registered users. Sign in to your account and get it there.
On the account page you get two options:
— Download installer, if the browser is on the machine you are installing.
— Install on a server, which gives you a one-line command to paste over SSH. That is the usual case: the browser is on your laptop, the server is somewhere else.
Linux users only — run this one line first (else the search engine won't start). Mac users skip it:
$ sudo sysctl -w vm.max_map_count=262144
Pasted the one-line command from your account page? It already unpacks and runs the installer — wait for it to finish, then go to step 3. Downloaded the file instead? Unpack it and run the installer: three commands, the rest is automatic. Grab a coffee for 10–20 min; it tells you how to open it when done.
$ tar xzf sovnex-install*.tar.gz -C ~ $ cd ~/sovnex-install $ ./install.sh ==> checking… generating secrets… pulling… starting… ✅ Sovnex is up.
Unpack on the server's own disk. Do not run it from a USB stick, and do not unpack on Windows and upload the folder: both strip the permission that lets install.sh run, and you get Permission denied.
Every image is public — there is no registry to log in to. Everything is automatic — no config to edit.
Large NVIDIA servers (vLLM): the model is about 31 GB. The installer finishes before it has loaded, and chat answers only once every model server shows (healthy) in docker ps --filter name=sovnex-vllm. Follow the download with docker logs -f sovnex-vllm-chat.
Write down the admin login it prints at the end — admin@sovnex.io and a password generated for this install (also in .env as DEFAULT_SUPERUSER_PASSWORD). There is no factory default password any more.
Open the Admin console in your browser and sign in with the login the installer printed → License. Click Purchase a plan and pay: your deployment activates by itself in about 30 seconds. Already have an activation code from your vendor? Paste it under Activate online instead.
# Admin console (replace localhost with server IP) http://localhost/admin → License
No model setup needed: the installer has registered its AI models for every new user, with the right context size — nothing to add under Model providers. The Getting started guide covers the optional next step, routing models through the built-in API gateway.
NVIDIA servers running vLLM also get a rerank model, started and registered by the installer: it finds answers the wording missed — worth it when your questions and your documents are in different languages. Do not want it? Install with ./install.sh --no-rerank.
<host> = localhost on this machine; use the server's IP or domain for remote access.
./install.sh down./install.sh up./install.sh status./install.sh logs./install.sh upgrade./install.sh logs