~/ / AI articles / Local LLMs on a VPS / Start a local LLM without an expensive PC: a beginner's guide to using a VPS
Start a local LLM without an expensive PC: a beginner's guide to using a VPS
You do not need to buy a powerful computer to try a local language model. You can rent a server, called a VPS, install the model there, and talk to it from the browser of any device. This guide takes you from nothing to a working, protected chat screen in six steps. It uses the tools of this site along the way, and it says plainly what a VPS without a GPU can and cannot promise.
Published: · 5 min read · Facts checked against the official documentation and the official pages of the VPS services on 2026-10-26. Plans, prices and defaults change, so confirm them before you pay.
What you can expect, and what you cannot
A VPS is a computer in a data center that you rent. A local LLM here means a model that runs on that computer, instead of a service that you call over the internet. Your chats stay on a server that you control.
A VPS is not a replacement for a PC with a strong graphics card, and this guide does not claim that. This is what we could confirm, and what we could not.
| Question | What we found |
|---|---|
| Do the VPS plans have a GPU? | None of the six VPS pages that we checked lists a GPU plan. The pages of ConoHa and XServer point to separate GPU services. |
| Then where does the model run? | In the server's memory, on the CPU. The Ollama FAQ says that ollama ps shows 100% CPU when a model was loaded entirely in system memory. |
| How fast is it? | We do not know. The documents we checked give no speed figures for CPU-only operation, and we did not measure any. Start with a small model and test it. |
| What decides whether it runs at all? | Memory. A model that does not fit in the server's memory will not run well. Step 1 shows how to estimate it. |
If you only chat now and then, a paid AI service that you call over the internet may cost less than a server that runs all month. Step 3 includes a calculator so that you can compare with your own numbers.
Step 1: estimate the memory you need
A model needs memory for its file, for the cache that grows with the length of the conversation, and for the operating system. The table shows our estimate for six models at a context length of 4,096 tokens, one request at a time, with 1 GiB reserved for the system. The estimates are calculated from the Ollama library sizes and the official model configurations, so they are estimates and not measurements.
| Model | Estimated total (GiB) | Smallest common memory size (GiB) |
|---|---|---|
| Qwen3 0.6B | 1.92 | 2 |
| Qwen3 1.7B | 2.74 | 4 |
| Qwen3 4B | 3.89 | 4 |
| Qwen3 8B | 6.41 | 8 |
| Qwen3 14B | 10.29 | 12 |
| Mistral 7B (v0.3) | 5.60 | 8 |
Try it: Local LLM Memory Estimator
Estimate
- Model weights
- 4.84 GiB
- KV cache
- 0.56 GiB
- Reserve
- 1.00 GiB
- Total
- 6.41 GiB (6.88 GB)
- Smallest common memory size that holds it
- 8 GiB
This is an estimate. It does not include the runtime's compute buffers, and it assumes that every layer has the same KV shape. Measure a real run before you choose a plan.
Model data checked on 2026-10-26.
Sources: Ollama library · Model configuration (config.json) · Ollama FAQ
Everything runs in your browser. Nothing is uploaded.
Step 2: choose a model
The Ollama library labels each model with what it can do. A model that carries tools can call functions that you define, thinking shows its reasoning separately from the answer, vision accepts images, and embedding turns text into vectors for search. For a first chat you need none of them, so pick by memory first.
| Tag | Labels | Download size | Memory class to plan for |
|---|---|---|---|
| qwen3:4b | tools, thinking | 2.5 GB | 4 GiB |
| llama3.2:3b | tools | 2.0 GB | not estimated here |
| gemma3:4b | vision | 3.4 GB | not estimated here |
| mistral:7b | tools | 4.4 GB | 8 GiB |
The download size is the least memory a model needs. Models without an estimate are not in the calculator's presets, because it needs their layers and head sizes from the model's configuration; you can type them in as a custom model.
Step 3: choose a VPS
Look at memory first, then at the disk for the model files, then at the CPU. The disk matters because one model file is several gigabytes. Then look at how much of the setup the service has prepared. Of the six services at the end of this guide, the pages of ConoHa VPS, KAGOYA CLOUD VPS and XServer VPS mention Dify, and the page of ConoHa VPS also lists n8n and Claude Code among its startup scripts. None of the six list a model server such as Ollama, so you install it yourself in Step 4.
The section at the end of this page compares the six services from their official pages. Here you can compare a server with an AI service that you call over the internet.
Try it: LLM API Cost Calculator
Estimated cost
Enter the prices and token counts to see the cost.
Official pricing pages
This only multiplies the numbers you enter. It does not know the vendor's prices, discounts, caching, batch rates, tiered pricing or tax.
To find the input tokens of a prompt, use the Token Counter.
Step 4: install Ollama
These are the commands from the official Ollama documentation for Linux. Read a script before you run it on a server that matters.
| Step | Command |
|---|---|
| Install | curl -fsSL https://ollama.com/install.sh | sh |
| Check that it runs | ollama -v |
| Run a model from the library | ollama run qwen3:4b |
To start Ollama at boot and keep it running, the documentation describes a systemd service. The full steps, the settings and the updates are in the detailed articles listed near the end.
Step 5: chat in your browser
Typing in a terminal is enough to test a model. For everyday use, Open WebUI gives you a chat screen in the browser. Its quick start runs it with one command and serves it at http://localhost:3000. Replace your-secret-key with a value that you generate with openssl rand -hex 32.
| Step | Command |
|---|---|
| Start Open WebUI | docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data -e WEBUI_SECRET_KEY=your-secret-key --name open-webui --restart always ghcr.io/open-webui/open-webui:main |
- The first account that you create is the administrator, and sign-up switches itself off once the admin account exists. Create it right away.
- The guide shows two ways to connect Open WebUI to Ollama, and the detailed article explains both.
Step 6: keep it private
Ollama listens on 127.0.0.1 port 11434 by default, which is the safest setting, and its documentation describes no authentication of its own. Our recommendation, not a statement from the documents: do not open that port to the internet. Put nginx in front with a password and HTTPS, and use a firewall that admits only the addresses you use.
- Keep the secret key and the passwords out of anything you commit.
- Update Ollama and Open WebUI from time to time, and keep the Open WebUI data volume when you do.
Go deeper
- 1How much memory does a local LLM need on a VPS? Model file plus KV cacheModel file plus KV cache: what the Ollama library, the Ollama FAQ and the official model configurations say about the memory a local LLM needs on a VPS.
- 2What can a local model do? Tools, thinking, vision and embeddings in Ollama, and what fits on a VPSTools, thinking, vision, embeddings: what the labels in the Ollama library mean, which models carry them, and how big the downloads are.
- 3Install Ollama on a VPS and keep it running: install, systemd service, settings and updatesInstall Ollama on a Linux VPS and keep it running: the install command, the systemd service, environment variables, model storage and updates, from the official docs.
- 4Open WebUI on a VPS: run it in Docker and connect it to OllamaRun Open WebUI with Docker and connect it to Ollama on your own server: the commands, the first account, the two connection types and how to update.
- 5Put nginx in front of Ollama: basic authentication, HTTPS and streaming settingsReach Ollama from outside safely: nginx as a proxy, basic authentication, HTTPS, streaming and timeout settings, and a firewall rule, from the official docs.
- 6Running AI on your own VPS: how much memory to plan for, and what not to exposePlanning to run Ollama or Dify on a VPS? What the official docs say about memory, CPU-only inference and the ports to keep closed.
Sources
Read next
- Put nginx in front of Ollama: basic authentication, HTTPS and streaming settingsReach Ollama from outside safely: nginx as a proxy, basic authentication, HTTPS, streaming and timeout settings, and a firewall rule, from the official docs.
- Open WebUI on a VPS: run it in Docker and connect it to OllamaRun Open WebUI with Docker and connect it to Ollama on your own server: the commands, the first account, the two connection types and how to update.
- Calling a local model from your code: structured outputs and the OpenAI-compatible API in OllamaConstrain a local model's output with a JSON Schema and call Ollama from OpenAI-style clients: what the official docs say, and what is not supported.
- What can a local model do? Tools, thinking, vision and embeddings in Ollama, and what fits on a VPSTools, thinking, vision, embeddings: what the labels in the Ollama library mean, which models carry them, and how big the downloads are.