CognenzaAI
EN
Tools · Level intermediate

Ollama: how to use AI models on your computer, offline

A beginner’s guide to Ollama: what it is, memory requirements, installation on Windows, Mac and Linux, first model, local-only mode, privacy and limits, from the official documentation.

By Cognenza editorial team · Updated · 4 min read · Sources cited

What it is

Ollama is an open-source program that downloads open-weight models and runs them on your own computer. Once the model is downloaded, you can chat offline without your texts leaving your computer. We introduced it in the radar; this guide explains how to get started.

Requirements

  • System: macOS, Windows or Linux.
  • Memory: depends on the model. For the official guide’s local example (Gemma 4 E2B, a download of about 7.2 GB), 8 GB of video memory (VRAM), or unified memory on a Mac, is recommended. With less, Ollama uses system RAM and responds more slowly. Larger context windows need more memory.
  • Disk: space for each model you download.

Step-by-step installation

  1. Download Ollama from ollama.com/download and install it.
  2. Open the app and choose a local model, or open a terminal and type:
ollama run gemma4:e2b
  1. Ollama downloads the model and opens a chat. Type your question; to exit, /bye.
  2. To see which models are loaded and whether they use the graphics card: ollama ps.

Privacy: local models vs the cloud

Ollama also offers cloud models (with an account). According to its documentation, with local models it doesn’t see your prompts or data. If you want to make sure everything stays on your computer, you can turn off cloud features: we explain how in the Ollama local-only recipe.

Precautions

  • By default Ollama only listens on your own computer (127.0.0.1:11434). Don’t open it to your network or the internet unless you know how to protect it: anyone with access could use it.
  • Download models from known sources and check their licence.
  • A local model can also make things up (hallucinations): verify what matters.

Limits

  • Models that fit on a home computer tend to be less capable than the big cloud assistants.
  • Without a powerful graphics card, responses can be slow.
  • The default context is 4096 tokens; for long documents you need to increase it (see the recipe).

Frequently asked questions

Does Ollama send my conversations over the internet?

According to its documentation, when you use local models Ollama doesn’t see your prompts or data. If you use its cloud models, it processes your prompts and responses to provide the service, but says it doesn’t store or log them and never trains on them.

Which model should I try first?

A small one. The official guide uses Gemma 4 E2B as its local example: it takes about 7.2 GB and recommends 8 GB of video memory. With less, it can use system RAM, though it will be slower.

Is a local model as good as ChatGPT?

Usually not: models that fit on a home computer tend to be less capable than the big cloud assistants. In return, your data doesn’t leave your computer.

Can I use it with other apps?

Yes. Ollama provides a local API (on port 11434) used by many apps, such as Open WebUI, OpenBot or Hermes Agent.

Keep reading

Sources

Official sources are linked in their original language.

  1. Quickstart (documentación oficial) — Ollama
  2. FAQ (documentación oficial) — Ollama
  3. Descargas de Ollama — Ollama
  4. Repositorio oficial ollama/ollama — GitHub

Change log

  • : First published with the official documentation and version 0.40.5 (14 October 2026).

Educational content, prepared with the help of AI tools, with sources cited. It is not a substitute for professional advice. Tools change fast: if you spot something out of date, write to us at contacto@cognenza.com. How we work.