Tutorial, Ubuntu 24.04

Ollama on a VPS. Your own model API, on CPU or GPU, kept private.

Most Ollama VPS guides end with OLLAMA_HOST=0.0.0.0 and an open port. That publishes a free model API for anyone who scans for it. Here’s the setup that stays yours.

10 min read
5 chapters
Updated 3 November 2026

What Ollama's docs say

Listens on127.0.0.1, port 11434Local only
Local API loginRequests are not checkedNone
Context, small GPUsDefault length4,096

So keep it local, and put a key in front if you open it.

Ollama on a Lite 16

$20

16 GB RAM
Qwen3 14B
Llama 3.1
Gemma 3
DeepSeek-R1
Mistral

Fits 14B models on CPU by our sizing rule. Billed by the hour.

The short answer

Install with Ollama’s one-line script, pull a model, and call its API on port 11434 of the server itself. It runs on CPU with no GPU at all; the model has to fit in RAM, so a 16 GB server runs models up to about 14B.

Ollama doesn’t check local API requests, so never open port 11434 to the internet. Use an SSH tunnel, or a proxy that checks a key: the builder below writes both.

Ollama downloads open models and serves them through an API, its own and an OpenAI-compatible one. On a VPS it becomes a model endpoint for your apps, agents and scripts that runs all day.

The two things to get right are memory, because the model lives in RAM or GPU memory, and access, because the API has no login of its own.

In this guide you’ll learn

CPU or GPU, and which models fit
Install and check it’s running
Call the API, native and OpenAI-style
A settings builder for private remote access
Fixes for the common errors

Chapter 1

CPU or GPU: which server for Ollama?

Ollama runs on CPU when there’s no GPU, using system RAM. A GPU is faster, but only for what fits in its memory: a bigger model splits between GPU and RAM, and ollama ps shows the split.

What runs fully, by our sizing rule (model file size plus cache, with 2 GB kept for the system)
Server	Memory	Biggest that runs fully	Price
Lite 8	8 GB RAM	8B models (Llama 3.1 8B, Qwen3 8B)	$12 a month
Lite 16	16 GB RAM	14B models (Qwen3 14B, Phi-4)	$20 a month
Lite 32	32 GB RAM	30B (Qwen3 30B, Gemma 3 27B)	$40 a month
Forge L40S	8 GB GPU, 32 GB RAM	8B on the GPU; bigger ones split	$400 a month

Every model and plan is in our model fit tool. We don’t publish speed figures we haven’t measured. Measure yours in one command: ollama run llama3.2:3b --verbose prints an “eval rate” in tokens a second.

Chapter 2

Install Ollama on Ubuntu

The official script installs Ollama, creates an ollama user and a systemd service, and starts it. With no GPU it says so and runs in CPU mode.

Install

curl -fsSL https://ollama.com/install.sh | sh

Check it runs

systemctl status ollama
ollama --version

Pull and try a model

ollama pull llama3.2:3b
ollama run llama3.2:3b "Say hello in five words"

Models are stored in /usr/share/ollama/.ollama/models. Logs: journalctl -e -u ollama. To update, run the install script again.

Chapter 3

Call the Ollama API

Ollama serves two APIs on port 11434: its own under /api, and an OpenAI-compatible one under /v1, so most OpenAI clients work by changing the base URL.

01
List your models
```
curl http://127.0.0.1:11434/api/tags
```

Chat, the Ollama way

curl http://127.0.0.1:11434/api/chat -d '{"model": "llama3.2:3b", "messages": [{"role": "user", "content": "Hello"}], "stream": false}'

Chat, the OpenAI way

curl http://127.0.0.1:11434/v1/chat/completions -H "Content-Type: application/json" \
  -d '{"model": "llama3.2:3b", "messages": [{"role": "user", "content": "Hello"}]}'

04
See what is loaded
```
ollama ps
```
It shows each loaded model and where it runs: 100% CPU, 100% GPU, or a split.

Chapter 4

Settings and private remote access

Ollama listens on 127.0.0.1 by default, and its docs say local requests need no authentication. Keep it that way, and choose how you reach it: an SSH tunnel for you alone, or a proxy that checks a key for apps elsewhere.

Your setupReach it fromRequests at once, per modelContext length (tokens)Keep a model loaded for

1. sudo systemctl edit ollama, paste, save

[Service]
Environment="OLLAMA_HOST=127.0.0.1:11434"
Environment="OLLAMA_NUM_PARALLEL=1"
Environment="OLLAMA_CONTEXT_LENGTH=4096"
Environment="OLLAMA_KEEP_ALIVE=5m"

sudo systemctl daemon-reload && sudo systemctl restart ollama

2. Connect from your computer

ssh -N -L 11434:127.0.0.1:11434 you@your-server-ip
# then, on your computer:
curl http://localhost:11434/api/tags

Deploy a Lite 16

For the proxy, install Caddy (sudo apt install -y caddy), point your domain at the server, paste the Caddyfile and run sudo systemctl reload caddy. Caddy gets the HTTPS certificate itself. Then open only ports 22, 80 and 443 in your firewall.

Chapter 5

Common Ollama errors on a server

The problems people hit most, and what to check.

Killed, or out of memoryHow to check: The model plus its cache didn’t fit. Use a smaller model, lower the context, or set OLLAMA_MAX_LOADED_MODELS=1 so switching models doesn’t keep two in memory. Adding swap stops the crash but is slow.
100% CPU on a GPU serverHow to check: Check nvidia-smi first. If the GPU is fine, the model is bigger than GPU memory: ollama ps shows the split. Pick a smaller model.
403 behind a proxyHow to check: Some setups need the proxy to send Host 127.0.0.1:11434; the Caddyfile above does.
Slow answersHow to check: On CPU, smaller models answer faster. Measure with --verbose, then try a 3B or 4B model before buying a bigger server.

MythOllama needs a GPU.

Fact

It runs on CPU with system RAM; the installer says so when there’s no GPU.

MythSetting an API key in the client secures Ollama.

Fact

Ollama ignores it. Only the network setup secures it.

Now it’s your move

Install with one script, pick a model that fits your memory, keep the API on 127.0.0.1, and reach it through a tunnel or a key-checking proxy.

Want a chat interface on top? A private AI server adds Open WebUI in one click. Serving many users at once on a GPU? Compare vLLM and Ollama.