LALocal AI Stack

Guide

How to Install Ollama on Mac, Windows, and Linux

Install Ollama on Mac, Windows, or Linux, run your first local AI model, verify that it works, and avoid common storage, PATH, and first-run mistakes.

Verdict

Official documentation reviewed, with caveats

Evidence label: Official documentation reviewed, with caveats. Sources were reviewed on 2026-05-24. Local AI Guide test status: Not independently tested by Local AI Guide. This page does not contain local benchmark, install, privacy-audit, network-monitoring, storage-inspection, or screenshot evidence.

Ollama is one of the simplest ways to run local AI models from your own computer. Install the app for your operating system, open a terminal, run a small model first, and then decide whether you want to add a browser interface like Open WebUI later.

The fastest beginner path is:

  1. Install Ollama for your operating system.
  2. Open Terminal, PowerShell, or Command Prompt.
  3. Run a small starter model.
  4. Confirm that Ollama responds locally.
  5. Only then move to larger models, Open WebUI, document chat, or custom integrations.

Quick recommendation

Use Ollama if you want a lightweight local model runner, terminal workflow, local API access, or a backend for tools like Open WebUI. Use LM Studio instead if you want a beginner-friendly desktop app for finding, downloading, and chatting with local models through a graphical interface.

If you are new to local AI, do not start by downloading the largest model you can find. Start with a small model, confirm that the setup works, and then move up based on your RAM, VRAM, and patience.

Before you install Ollama

Ollama itself is not usually the hard part. The bigger issues are model size, storage, RAM, VRAM, and expectations.

QuestionBeginner answerEvidence label
Do I need a GPU?No for very small models, but a GPU or Apple Silicon unified memory usually makes the experience much better.Conservative estimate, not a benchmark
Can I use an 8GB laptop?Yes for small models and modest expectations. Do not expect large models or heavy PDF workflows to feel smooth.Conservative estimate, not a benchmark
Is 16GB RAM enough?Usually yes for a practical beginner setup with small or 7B/8B-class models.Conservative estimate, not a benchmark
Does Ollama include a polished chat app?Ollama is primarily a model runner, CLI, and local API. Add Open WebUI or another interface if you want a browser-based chat app.Official documentation reviewed
Does local mean private?Local inference can keep prompts on your machine, but model downloads, cloud features, web search, exposed local servers, and third-party tools can change the privacy profile.Official documentation reviewed + Conservative estimate, not a benchmark

System requirements and storage basics

Use this table as a practical starting point, not as a guarantee that every model will run well.

PlatformPractical beginner notesEvidence label
MacPrefer Apple Silicon for the best beginner experience. Keep enough free disk space for model files.Official documentation reviewed + Conservative estimate, not a benchmark
WindowsOllama runs as a native Windows application and supports NVIDIA and AMD Radeon GPU paths. Windows 10 22H2 or newer is the official Windows requirement.Official documentation reviewed
LinuxOllama provides a Linux install path and can be used as a service-style local runtime. Distribution details vary.Official documentation reviewed
StorageThe app is only the start. Downloaded models can take many gigabytes, and a model library can grow quickly.Official documentation reviewed + Conservative estimate, not a benchmark
RAM/VRAMModel size, quantization, and context length determine whether the setup feels usable.Conservative estimate, not a benchmark

For Windows, Ollama’s documentation states that the binary install needs at least 4GB of space, with additional space required for large language models that can range from tens to hundreds of gigabytes depending on what you download.

Which first model should you run?

For a beginner install guide, the first model should be small enough to prove that the setup works. Your first model is not your forever model.

Your machineFirst model class to tryWhat to avoid at firstEvidence label
8GB system RAMSmall 3B/4B-class modelLarge 13B/14B+ models, long context, multiple apps openConservative estimate, not a benchmark
16GB system RAMSmall model or 7B/8B-class modelAssuming every 14B model will be comfortableConservative estimate, not a benchmark
32GB system RAM7B/8B comfortably; some larger models may be practicalJumping straight to 70B-class modelsConservative estimate, not a benchmark
CPU-only machineSmall model with modest expectationsTreating CPU-only as equivalent to GPU accelerationConservative estimate, not a benchmark
Windows with NVIDIA GPUSmall model first, then test larger models based on dedicated VRAMPlanning around shared GPU memory as if it were dedicated VRAMConservative estimate, not a benchmark
Apple Silicon MacSmall model first; move up based on unified memoryForgetting that macOS and apps share the same memory poolConservative estimate, not a benchmark

Local AI Stack’s testing protocol uses small and medium baseline model classes for repeatability. For an actual hands-on article, keep the exact model identifier constant across screenshots and measurements.

Install Ollama on Mac

1. Download Ollama

Go to the official Ollama download page and choose the macOS download.

Evidence label: Official documentation reviewed.

2. Open the app

Open the downloaded app and complete the normal macOS install flow. If macOS asks for permission to open the app, follow the standard macOS security prompt.

Evidence label: Official documentation reviewed.

3. Open Terminal

After the app is installed and running, open Terminal. The command line is where you will run your first model.

4. Check that the command works

Run:

ollama --version

If the command returns a version, the CLI is available.

If the command is not found, quit and reopen Terminal. If it still fails, reopen the Ollama app and check the official macOS setup notes.

5. Run a small first model

Run a small model first. For example:

ollama run gemma3:4b

Then ask a simple test question:

Explain local AI in one sentence.

If the model replies, your basic install works.

Publisher note: Before final publication, verify that the starter model name is still available in the Ollama library. Replace it with your house baseline if Local AI Stack standardizes on a different model.

Install Ollama on Windows

1. Confirm your Windows version

Ollama’s Windows documentation lists Windows 10 22H2 or newer as the Windows requirement. If you are using an older Windows build, update before installing.

Evidence label: Official documentation reviewed.

2. Download the Windows installer

Download the official Windows installer from Ollama.

3. Run the installer

Ollama’s Windows documentation says the installer does not require Administrator rights and installs in your home directory by default.

Evidence label: Official documentation reviewed.

4. Open PowerShell, Command Prompt, or Windows Terminal

After installation, Ollama should run in the background, and the ollama command should be available in cmd, PowerShell, or your preferred terminal.

Run:

ollama --version

If the command is recognized, continue.

5. Run your first model

Start with a small model:

ollama run gemma3:4b

If this is your first local model download, expect the first run to take longer because Ollama has to download the model before it can run it.

6. Verify the local API if you need it

Ollama’s Windows documentation states that its API is served at:

http://localhost:11434

For most beginners, you do not need to use the API immediately. But if you are setting up Open WebUI or another local tool, this local endpoint is important.

A simple PowerShell test looks like this:

(Invoke-WebRequest -Method POST -Body '{"model":"gemma3:4b", "prompt":"Say hello in one sentence.", "stream": false}' -Uri http://localhost:11434/api/generate).Content

If you receive a JSON response, the local API is responding.

Install Ollama on Linux

1. Use the official Linux install path

Follow Ollama’s official Linux installation documentation. The exact steps may vary depending on whether you want a normal local setup, service-style setup, GPU support, or Docker.

Evidence label: Official documentation reviewed.

2. Confirm the command is available

Run:

ollama --version

3. Run a small first model

Run:

ollama run gemma3:4b

Then ask a short test prompt.

4. Check whether the service is running

If ollama run fails because the local server is not available, check the official Linux troubleshooting and service documentation. Linux users are more likely than Mac or Windows beginners to encounter service, permission, or GPU-driver issues.

Do you need Docker for Ollama?

Most beginners do not need Docker just to install and use Ollama.

Use the normal app or OS install path first. Consider Docker only if:

  • you already use Docker;
  • you want an isolated deployment;
  • you are following a specific Open WebUI setup that expects containers;
  • you are deploying on a server;
  • you need repeatable infrastructure rather than a normal desktop install.

If your goal is simply “run my first local AI model,” Docker is usually an unnecessary complication.

Where Ollama stores files

Model storage matters because the app may be small, but the models are not.

ItemWindows location from Ollama docsMac/Linux practical noteEvidence label
Logs and updates%LOCALAPPDATA%\OllamaCheck official docs for current pathsOfficial documentation reviewed for Windows
Binaries%LOCALAPPDATA%\Programs\OllamaApp or binary path depends on install methodOfficial documentation reviewed for Windows
Models and config%HOMEPATH%\.ollamaCommonly under ~/.ollamaOfficial documentation reviewed for Windows; verify for Mac/Linux before publication
Temporary files%TEMP% with ollama* foldersTemporary paths varyOfficial documentation reviewed for Windows

If your system drive is small, solve storage before downloading several models.

How to move Ollama models to another drive on Windows

Ollama’s Windows documentation says you can change the model location by setting the OLLAMA_MODELS environment variable for your user account.

Beginner-safe outline:

  1. Create a folder on the target drive, such as D:\OllamaModels.
  2. Open Windows Settings or Control Panel.
  3. Search for environment variables.
  4. Edit environment variables for your user account.
  5. Create or edit OLLAMA_MODELS.
  6. Set it to the folder where you want downloaded models stored.
  7. Quit the Ollama tray app.
  8. Relaunch Ollama from the Start menu.
  9. Open a new terminal and run a small model.

Important: If you already downloaded models, moving the variable does not automatically move old model files. Back up before deleting anything.

How to verify Ollama is working

Use these checks in order.

CheckCommand or actionSuccess looks likeEvidence label
CLI existsollama --versionVersion output appearsOfficial documentation reviewed
Model list worksollama listA table appears, even if emptyOfficial documentation reviewed
Model runsollama run gemma3:4bModel downloads, loads, and repliesOfficial documentation reviewed + Conservative estimate, not a benchmark
Local API respondsPOST to http://localhost:11434/api/generateJSON response appearsOfficial documentation reviewed
Storage is understoodCheck model folderYou know where downloaded files liveOfficial documentation reviewed for Windows; verify others

Common Ollama install problems

ProblemLikely causeFixEvidence label
ollama command not recognizedTerminal was opened before install, PATH not refreshed, or app not launchedClose and reopen the terminal; relaunch Ollama; check the official install notesConservative estimate, not a benchmark
Windows install uses the wrong driveDefault home-directory install pathStart the installer with a different directory flag or use documented storage settingsOfficial documentation reviewed
Models fill the C driveDownloaded models are much larger than the appSet OLLAMA_MODELS before building a model libraryOfficial documentation reviewed
First run takes a long timeModel download is happening before inferenceWait for the download; start with a small modelConservative estimate, not a benchmark
Responses are painfully slowModel is too large, CPU-only fallback, not enough memory, or long contextTry a smaller model and close other memory-heavy appsConservative estimate, not a benchmark
Progress symbols look strange on Windows 10Terminal font may not render Ollama progress characters correctlyChange terminal font settingsOfficial documentation reviewed
Local API does not respondOllama background process is not runningStart Ollama, then retry localhost:11434Conservative estimate, not a benchmark
Open WebUI cannot see OllamaContainer/networking issue or Ollama not reachable from the UIFinish this install first, then use the Open WebUI with Ollama guideConservative estimate, not a benchmark
User assumes it is fully privateLocal model execution is only one part of the workflowRead the privacy caveat and avoid cloud-connected features when handling sensitive dataOfficial documentation reviewed + Conservative estimate, not a benchmark

Privacy notes for Ollama

Ollama can run models locally, and the local API is served from your own machine. But do not treat installation alone as a complete privacy plan.

A local setup can still involve network activity when you:

  • download models;
  • update the app;
  • use hosted cloud models;
  • enable web search or other cloud features;
  • connect third-party tools;
  • expose the local server beyond your own machine;
  • use tunnels, proxies, or remote access;
  • pair Ollama with an interface that connects to cloud APIs.

The safe beginner assumption is:

A model run locally through Ollama is more private than sending the same prompt to a cloud chatbot, but only the parts of your workflow that actually stay on your machine are local.

Do not expose localhost:11434 to the public internet unless you understand authentication, network binding, firewall rules, and the security consequences.

Should you install Open WebUI next?

Install Open WebUI after Ollama if you want a browser-based chat interface, multi-user-style UI, document workflows, or a more ChatGPT-like front end.

Do not install it immediately if you are still unsure whether Ollama works. First confirm:

  1. ollama --version works.
  2. ollama run [model] works.
  3. You know where models are stored.
  4. You have enough disk space.
  5. You understand that a local UI can still be configured to use cloud providers.

Should you install LM Studio instead?

Install LM Studio instead of Ollama if you want:

  • a graphical model browser;
  • a desktop chat interface;
  • easier first-model discovery;
  • built-in document chat;
  • a local server only when you need it;
  • fewer terminal commands at the start.

Use Ollama if you want:

  • a lightweight local runtime;
  • command-line model management;
  • a local API backend;
  • Open WebUI integration;
  • developer tooling;
  • automation workflows.

Sources and evidence

Official and research sources used for this draft:

  • Ollama Quickstart: https://docs.ollama.com/quickstart
  • Ollama macOS installation: https://docs.ollama.com/macos
  • Ollama Windows documentation: https://docs.ollama.com/windows
  • Ollama Linux documentation: https://docs.ollama.com/linux
  • Ollama Docker documentation: https://docs.ollama.com/docker
  • Ollama API documentation: https://docs.ollama.com/api
  • Ollama privacy page: https://ollama.com/privacy
  • Local AI Stack keyword map and SERP research packet
  • Local AI Stack source-of-truth research packet
  • Local AI Stack compatibility foundation
  • Local AI Stack privacy and security research packet
  • Local AI Stack repeatable testing protocol

Fact status

Official documentation reviewedNot independently tested by Local AI GuideReviewed: 2026-05-24
  • Local AI Guide has not independently installed, benchmarked, or audited this workflow.
  • Follow official documentation for current commands, requirements, provider settings, and privacy boundaries.