Install Ollama on Windows, Linux, and Mac

By Emmanuel Sauvage · Published · Updated

This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.

Ollama is a free app that downloads and runs LLMs (large language models) on your computer. The weights are often served as GGUF, a format compressed by quantization (Q4, Q5, and so on) that cuts disk size and memory needs. VRAM (video RAM, the graphics card’s memory) or system RAM holds the model. A lighter quantization needs less memory, and it also gives up some quality.

Plan on 25 to 40 minutes, depending on your system and on the speed of the first model download (network troubleshooting is extra). This guide covers Windows, Linux, and macOS with concrete commands, fixes for common errors (systemd, the firewall, PATH), and pointers to a first local AI in 15 minutes, choosing a first model, and Open WebUI for a family interface.

Commands to copy as written — Use only the text blocks (fenced blocks) in this guide: one command per line, and an Ollama tag with no spaces (for example mistral:7b-instruct-q4_K_M). The tables, including the GGUF quantization table, are for reading. Do not copy from a cell. The layout can insert spaces that make the tag invalid.

The first command that should work (Windows, Linux, or macOS)

Once Ollama is installed on your OS (the steps below), run these three lines in order in PowerShell, Terminal, or bash:

ollama --version
ollama pull llama3.2:3b
ollama run llama3.2:3b

What you should see: a version number (never “command not found”), a download that ends in “success”, then a reply to a short test question. Type /bye to leave the chat. If ollama --version fails, fix PATH before you continue. A more detailed Windows guide (GPU, instruct tags): install Ollama on Windows.

The Windows, Linux, and macOS sections send you back here after the install. Do not copy a second command block somewhere else.

Which page should you read?

Goal Where to start Rough time Who
The shortest Windows path 15 minutes 12–18 min A complete beginner, one OS
Windows, step by step Install Ollama on Windows about 25 min (depends on download speed) An NVIDIA GPU, PATH troubleshooting
Several operating systems (this guide) The OS filter, plus the Windows, Linux, or macOS section 25–40 min Mac, Linux, several machines, or an intermediate reader who is fine with a terminal
A container stack Docker for Ollama and Open WebUI only about 30 min You already know Docker. Use that path on its own, not beside a native install

Read only one OS section. Use the filter below or the contents list, confirm the first command box, then read GGUF quantization only if you change models.

Filter this guide — Pick your system to hide the install chapters for the other operating systems (Windows, Linux, macOS). Prerequisites, GGUF tags, native versus Docker, security, and what comes next stay visible.

With no filter active, follow a single link in the contents list (section 2, 3, or 4).

What you will have at the end

Piece Expected state
The Ollama binary ollama --version answers in the console
A first model llama3.2:3b (or the GGUF tag you chose) shows up after ollama list
A CLI chat At least one question and reply. Leave with /bye
GPU or CPU Task Manager, nvidia-smi, or Activity Monitor shows the load

Contents

Native or Docker — pick one path: install Ollama through Windows, Linux, or macOS (sections 2 to 4) or follow only the Docker guide for Ollama and Open WebUI. Running both at once conflicts on port 11434. Comparison: Native or Docker.

Common copy-and-paste mistakes (every OS)

Mistake Symptom Fix
A space inside ollama --version “command not found” Copy the text block from the box above or from your OS section
A tag with spaces or broken hyphens pull fails, unknown model Copy a whole line from the quantization block (for example mistral:7b-instruct-q4_K_M)
The terminal was not reopened after install ollama is not found Close the terminal and open it again. Check that the Ollama service is running
Native Ollama plus Docker Ollama Port 11434 is already in use One path only. See native or Docker

Before you install: resources and OS

Need Guidance
Disk Several gigabytes per model. Prefer an SSD for the models
RAM 16 GB minimum for small models. 32 GB is more comfortable — 32 vs 64 GB of RAM
GPU Optional, and useful. See VRAM and model size and GPU and eGPU

If ollama pull fails for lack of space, or the PC swaps heavily (it uses the disk as extra memory), a hardware upgrade can beat a recurring cloud subscription. How to choose: start with the hardware selector and choosing hardware for home AI (motherboard compatibility, where the drive goes). Amazon affiliate links (generic searches, October 2026): NVMe SSDs on amazon.co.uk, a 32 GB memory kit on amazon.co.uk. Check the current price on amazon.co.uk and the spec sheet before you buy. Disclosure: affiliate page.

Useful docs, depending on the case: Apple Silicon to understand Metal acceleration, and a Mac Mini M4 for local AI for a dedicated macOS profile. WSL2: only the optional note at the end of the Windows chapter, after the .exe installer has succeeded.

Windows (10 and 11): native install

A step-by-step version is also in install Ollama on Windows.

  1. Download the installer from ollama.com/download (Windows).
  2. Run the .exe. Accept the Ollama background service if it is offered.
  3. Close and reopen PowerShell or Terminal so PATH refreshes.
  4. Confirm with the three shared commands (version, pull llama3.2:3b, run).
  5. During generation, open Task Manager, then GPU performance, to see whether NVIDIA is in use.

Common Windows errors

  • ollama is not recognized: PATH was not updated. Restart the terminal or the session. Confirm the installer finished.
  • Firewall: allow Ollama on the private network only. Leave port 11434 off the internet at the router.
  • No GPU: update the NVIDIA drivers. Reboot after the install.
  • Disk space: models live under the user profile (the .ollama folder). Free space, or move the folder, using the steps in the Ollama FAQ.
  • A corporate proxy: HTTPS_PROXY may be required for pull.

Only if the native installer does not fit your case

WSL2 (Windows Subsystem for Linux) is for installing Ollama inside a Linux distro, with an NVIDIA GPU through the Windows drivers. That is Linux-side development without a dual boot. It is not the path for a household that already confirmed the first command box with the .exe. Docs: Microsoft WSL2. If you choose this route after you have tried the native install, follow the Linux section inside the WSL distro.

Linux: the official script and the systemd service

On native Windows? Skip this section (it is for bare-metal Linux, a NAS, or WSL2). Under WSL2, systemd may be missing depending on the version. See the errors below.

  1. Update the system and install curl if you need it:
sudo apt update && sudo apt install -y curl
  1. Run the official install script:
curl -fsSL https://ollama.com/install.sh | sh
  1. Check the binary:
ollama --version
  1. The script usually sets up a systemd service named ollama.service. Check it:
sudo systemctl status ollama
sudo systemctl enable ollama
sudo systemctl start ollama
  1. (NVIDIA GPU) Install your distro’s proprietary drivers, then test nvidia-smi.
  2. Confirm with the three shared commands.
  3. For the API from another machine on the LAN, set OLLAMA_HOST (see the docs), and only on a network you trust.

Common Linux errors: journalctl and ufw (an advanced home server)

The journalctl notes (systemd logs) and ufw notes (Uncomplicated Firewall) below are for an advanced home server or a headless machine. On a desktop with a graphical interface, the built-in firewall is often enough. Leave ufw alone unless you have a backup of the rules.

  • Failed to start ollama.service: get the detailed log with journalctl -u ollama -e. Common causes are a missing binary, permissions, or a conflict on port 11434.
  • systemd not found (a minimal container): start ollama serve by hand in a user session.
  • PATH: if ollama is not found, the binary may be in /usr/local/bin. Check with which ollama.
  • Firewall (ufw): leave 11434 closed by default. If you open it, limit it by IP.
  • SELinux or AppArmor: rare blocks on model paths. Read the logs.

A headless server (NAS, mini PC): cross-check a Synology NAS and Ollama and noise and power when a mini PC stays on before you leave it running.

macOS: Apple Silicon and Intel

An Intel Mac with 8 GB of unified memory — Before you install the .dmg, aim for llama3.2:3b (the same first model as the other operating systems in this guide), close the browser and heavy apps, and watch memory pressure in Activity Monitor (Applications → Utilities). Past a quantized 7B, expect disk swap. See choosing a first model. The steps below stay the same.

Apple Silicon (M1, M2, M3, M4) versus Intel: Apple Silicon Macs speed up Ollama through Metal (Apple’s graphics API) and are the strong recommendation. Intel Macs still work, and they are much slower on large GGUF files. Prefer small models, or M-series hardware (Mac Mini M4).

  1. Download the .dmg from ollama.com/download (choose Apple Silicon or Intel).
  2. Drag Ollama into Applications and launch the app.
  3. Grant the permissions it asks for (System Settings → Privacy, if macOS blocks it).
  4. Confirm with the three shared commands.
  5. On Apple Silicon, GPU and Neural Engine activity shows up in Activity Monitor. Ollama handles Metal. There is no CUDA setup.

Common macOS errors

  • The app is blocked: right-click, then Open, the first time, or allow the developer in Settings.
  • Not enough memory: the model is too heavy. Use a more aggressive quantization or a smaller model (choosing a first model).
  • A slow Intel Mac: expected for large GGUF files. Prefer a light model, or consider M-series hardware (Mac Mini M4).
  • PATH in the CLI: if ollama is missing in a non-interactive Terminal, launch the binary in /Applications/Ollama.app/Contents/Resources/ or use the documented symlink.

Pick a GGUF quantization (every OS)

Once Ollama is installed, the tag you pull sets the tradeoff among size, quality, and memory:

Quantization Approx. size (7B) Quality Rough RAM or VRAM Ollama tag (reference — copy the text block below)
Q4_K_M ~4.5 GB A good household compromise 8–12 GB mistral:7b-instruct-q4_K_M
Q5_K_M ~5.5 GB A bit more nuance 10–14 GB mistral:7b-instruct-q5_K_M
Q8_0 ~7 GB Close to the original 12–16 GB mistral:7b-instruct-q8_0

Commands to copy (one block — do not copy from the table):

ollama pull mistral:7b-instruct-q4_K_M
ollama pull mistral:7b-instruct-q5_K_M
ollama pull mistral:7b-instruct-q8_0

For a light first test (every OS):

ollama pull llama3.2:3b

Tag names are explained in choosing a first model.

A 3B or an 8B in Q4 on 16 GB of RAM is still workable on CPU. A comfortable 13B often wants 32 GB, or a GPU with enough VRAM. VRAM and model size is how to weigh size against quantization.

Native install or Docker (Ollama alone)

Path When to choose it Typical next step
Native install (installer, script, or DMG) You are starting out, you want one service, and you want less RAM spent on overhead Open WebUI via pip or a binary
Docker Several containers (Ollama, Open WebUI, and volumes) Docker guide for Ollama and Open WebUI
Both on the same machine Avoid this Port 11434 conflicts, and VRAM is used twice

The native or Docker warning is at the top of the contents. A detailed Docker versus pip table for the interface is in Open WebUI for beginners.

After the install: models, the API, and interfaces

  1. List local models: ollama list.
  2. Remove a model: ollama rm name:tag.
  3. Update Ollama with the installer or the script, then test a pull again.
  4. Test the local HTTP API (from the same PC). Copy the text block:
curl http://127.0.0.1:11434/api/tags

You should see a JSON list of models already downloaded. If the connection is refused, the service is not running (Windows: the taskbar icon; Linux: systemctl start ollama).

  1. Point Continue or Home Assistant at the local URL on port 11434 (the Ollama API), following their docs.
  2. Compare with LM Studio if you want a test GUI before you settle on Ollama as the API backend.
  3. A very short path before this multi-OS guide: a first local AI in 15 minutes.

For images (ComfyUI, FLUX), Ollama does not replace a diffusion stack. See ComfyUI and FLUX locally and images without a cloud.

Security and privacy at home

Port 11434 and the LAN: Ollama exposes a local API (port 11434). Leave that port off your internet router’s forwarding list unless you have a secure tunnel and sign-in. Split off guest Wi-Fi if services listen on 0.0.0.0. Conversations stay local as long as you do not install a cloud sync. Read privacy when AI stays fully local before you load sensitive documents into a private family assistant. Back up the models folder with the rest of your data (SSD storage).

Common questions (Windows, Linux, macOS)

Full answers are in the body below. The collapsible Common questions block at the end of the article repeats the four frontmatter entries for a quick read in an accordion. Useful links: Docker versus native · GPU · Whisper · n8n.

Which version of Windows is supported?

Windows 10 and 11, 64-bit, with recent updates. The official installer runs the background service. WSL2 is not required for the native install.

Which Linux distribution works with the official script?

The curl script mainly targets Ubuntu, Debian, and derivatives. On Arch, Fedora, or others, the binary or a community package is often a better fit. The systemd service works the same way.

Intel Mac or Apple Silicon?

Apple Silicon (M1 and later) is the strong choice: Metal speeds up inference. An Intel Mac still works, and it is slower. See Apple’s docs and the Mac Mini M4 for local AI guide for the hardware.

Does Ollama listen on the internet by default?

It exposes a local API (port 11434). Leave that port off your router’s forwarding list unless you have a firewall and sign-in. See LAN security at the end of the article.

Should I use Docker instead of the native install?

Use only the native versus Docker table: one “core Ollama” path per machine. For the full container stack, follow the Docker guide for Ollama and Open WebUI afterward, without installing native Ollama beside it (a conflict on port 11434).

How do I know whether the GPU is in use?

Windows: Task Manager → Performance → GPU, while ollama run is going. Linux: the nvidia-smi command (NVIDIA drivers required). macOS: Activity Monitor → GPU or Neural Engine on Apple Silicon. If only the CPU climbs, use a smaller model (choosing a first model) or update the drivers.

Can I install this on an old PC?

Yes, for quantized 3B to 7B models, with 16 to 32 GB of RAM and an SSD. Skip mechanical hard drives for multi-gigabyte pulls.

Does Ollama replace ChatGPT?

For a lot of household jobs (a summary, a draft, general questions), yes, locally, and hallucinations are still possible. Compare habits and subscriptions in Ollama versus ChatGPT over 3 years. Hardware is a separate question.

Whisper, or transcription with Ollama?

They are different tools. For offline voice, see offline French transcription with Whisper and first steps with Whisper.

Home automations?

n8n and a local LLM call the Ollama API. Keep a single Ollama server so you avoid VRAM fights with other GPU apps.

Conclusion

Windows: the .exe installer, then the three shared commands. Linux: the official curl script, the systemctl service, and the same check. macOS: the Apple Silicon or Intel .dmg, and the same check (Metal on an M-series chip). WSL2: only if the optional note in the Windows chapter applies, after the native installer failed or you chose not to use it. Docs: Microsoft WSL2.

If something fails, check PATH and the firewall before you reinstall. On a Linux server or a NAS, see journalctl and ufw. Next software step: Open WebUI for beginners. If swap keeps happening, look at hardware: choosing hardware for home AI and VRAM.

Common questions

Which version of Windows is supported?

Windows 10 and 11, 64-bit, with recent updates. The official installer runs the background service. WSL2 is not required for the native install.

Which Linux distribution works with the official script?

The curl script mainly targets Ubuntu, Debian, and derivatives. On Arch, Fedora, or others, the binary or a community package is often a better fit. The systemd service works the same way.

Intel Mac or Apple Silicon?

Apple Silicon (M1 and later) is the strong choice: Metal speeds up inference. An Intel Mac still works, and it is slower. See Apple's docs and the LocAIhome Mac Mini M4 hardware guide.

Does Ollama listen on the internet by default?

It exposes a local API on port 11434. Leave that port off your router's forwarding list unless you have a firewall and sign-in. See the security section at the end of the article.

Sources

Narrow it down with the hardware selector · French version of this section