Install Ollama on Windows, step by step
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 the simplest app for running an LLM (a large language model) on your Windows PC. It downloads GGUF models (quantized files made for local inference) and offers a local API for Open WebUI or Home Assistant. Rough total time: about 25 minutes (installer around 5 minutes plus the first model; the weight download varies with your connection, before any network or antivirus troubleshooting).
Copy the commands as they are. Use only the text blocks below: one line per command, tags with no spaces. Retyping a command from the paragraph around it can insert spaces that break the tag.
ollama --version
What you want to see: a version number, never “command not found.”
Which page should you read?
| Goal | Page | Rough time | Who it is for |
|---|---|---|---|
| A first timed chat | 15 minutes | 12–18 min | Absolute beginner, Windows |
| Windows, step by step (this guide) | This page | ~25 min (depends on the model download) | Windows 10/11, GPU or CPU, troubleshooting |
| Mac, Linux, or several operating systems | Ollama on every OS | ~20–30 min | More than one platform |
| A container stack | Docker Ollama + Open WebUI | ~30 min | Already comfortable with Docker |
On this page
- What you need
- Steps 1 to 4
- Copy-paste mistakes
- Troubleshooting
- Advanced LAN (optional)
- What to do next
What you will have at the end
| Piece | What you want to see |
|---|---|
| Ollama service | The icon is active, and the check at the top of the page answers in the console |
| First model | An instruct Q4_K_M tag shows up after list (step 3) |
| CLI chat | At least one question and answer in English. Leave with /bye |
| GPU or CPU | Task Manager or nvidia-smi shows where the model is loaded |
What you need
| Piece | Detail |
|---|---|
| System | Windows 10 22H2 or later, or Windows 11 (64-bit) |
| Account | A local administrator for the installer |
| NVIDIA GPU (optional) | Current drivers. They speed the models up |
| Disk | An SSD is the comfortable choice, 15 GB free at minimum — model storage |
| RAM | 16 GB is a comfortable minimum; 32 GB is easier — 32/64 GB of RAM |
| Network | Internet for the installer and the first model download (step 3). Chat can work offline after that |
Terminal to use: PowerShell or Windows Terminal. Skip the old Command Prompt if you are copying commands from this guide.
Copy-paste mistakes that show up often
| Mistake | Symptom | Fix |
|---|---|---|
| A space in the version command | “command not found” | Copy the text block in step 2 |
| A tag with spaces or broken hyphens | Pull fails, unknown model | Copy the text block in step 3 (continuous tag llama3.2:3b-instruct-q4_K_M) |
| Terminal not reopened after install | ollama not found | Close PowerShell and open it again. Confirm the Ollama service is running |
| Service stopped during the pull | connection refused | Start Ollama from the Start menu and wait for the taskbar icon |
Numbered steps
1. Download Ollama
- Open a browser and go to https://ollama.com/download (or
/download/windows). - Click Download for Windows.
- Save OllamaSetup.exe (the name can shift slightly by version).
A common miss: downloading from an unofficial mirror. Use ollama.com only, so the installer is the real one.
2. Install
- Double-click the installer.
- Accept the Ollama service in the background (an icon in the notification area).
- Close PowerShell or Windows Terminal, then open it again. You need that so PATH (the list of commands Windows recognizes) refreshes.
Confirm the CLI answers:
ollama --version
What you want to see: a version number (for example 0.5.x), not “command not found.” If the command fails, restart the Windows session, or confirm the Ollama service is started (Apps, then Services, or the taskbar icon).
3. Download a small first model
An instruct model (trained to follow instructions in English or French) is a better fit for questions than a “base” model. Pick one tag for steps 3 and 4. Choose a first model explains Q4_K_M quantization (the size and quality compromise).
- 16 GB RAM (the recommendation for this tutorial):
ollama pull llama3.2:3b-instruct-q4_K_M
- 32 GB RAM or more (a slightly richer alternative):
ollama pull mistral:7b-instruct-q4_K_M
The download can take 10 to 20 minutes, depending on your connection. Leave the window open until the pull prints “success.”
Copy-paste block (16 GB RAM — one tag for steps 3 and 4)
ollama pull llama3.2:3b-instruct-q4_K_M
ollama run llama3.2:3b-instruct-q4_K_M
Paste one command at a time into PowerShell. The tag llama3.2:3b-instruct-q4_K_M contains no spaces: a colon before 3b, hyphens between instruct and q4_K_M. More on tags: choose a first model.
4. Test a conversation and see whether GPU or CPU is working
Use the same tag as in step 3. If you pulled Llama 3.2 3B:
ollama run llama3.2:3b-instruct-q4_K_M
If you pulled Mistral 7B:
ollama run mistral:7b-instruct-q4_K_M
Type a question in English (for example, “Explain in three sentences what a local LLM is”), then /bye to quit.
While it is generating, see where the model is running:
- NVIDIA: in PowerShell, run nvidia-smi (text block below if you need it). VRAM (the graphics card’s own memory) should rise if the GPU is in use.
nvidia-smi
- Task Manager, Performance, GPU: watch usage during the step 4 run session.
- If only the CPU climbs, the model is running on CPU. That is normal with no compatible GPU, or with a model that is too large. Cross-check VRAM and models before you buy a card.
What you want to see: the Ollama service is active (taskbar icon), the version block from step 2 answers, and the step 4 run session produces a coherent English reply with no “connection refused” error.
Steps 1 to 4 check out? Continue with what to do next below (the links are gathered in one place).
Troubleshooting
Hyper-V or virtualization conflict
Symptom: the installer stalls, or a message mentions virtualization. What to try: the Ollama Windows docs describe the native install, and WSL is not required. If you also run Docker Desktop with WSL2, leave virtualization enabled in the BIOS. If it still fails, check the official Ollama FAQ and try a session with Docker stopped. If you already know containers, switch to Docker Ollama + Open WebUI.
Not enough disk space
Models weigh several gigabytes. Move the storage or free the SSD — LLM model storage. On Windows, the default folder lives under %USERPROFILE%\.ollama.
A company proxy
Set the Windows system proxy variables, or download from a home network. The model download (step 3) goes out over HTTPS to the Ollama servers.
Antivirus and firewall
If the step 3 pull fails with a timeout or “forbidden,” allow OllamaSetup.exe and the Ollama service in Windows Defender (firewall and real-time protection). At work, ask for an exception for ollama.com (HTTPS port 443).
More than one graphics chip
On a laptop with an iGPU plus NVIDIA, Ollama should use the dedicated card when the Studio or Game Ready drivers are installed. If only the iGPU is working, update the drivers from NVIDIA and restart. Hardware detail is in GPU, iGPU, and eGPU.
Command not found after installation
If the version block from step 2 still errors after you restart the terminal: confirm the Ollama service is running (taskbar icon), restart the Windows session once, then test again in PowerShell (not the old cmd prompt). As a last resort, reinstall from ollama.com/download/windows.
Advanced LAN note (after the first success)
This section is optional. Finish steps 1 to 4 first, and set up Open WebUI if you want a family interface. The variables below are mainly for moving models onto a second disk, or for letting other machines on the LAN reach Ollama. The network side is covered in privacy when AI stays home.
| Variable | What it does |
|---|---|
OLLAMA_MODELS |
Folder for the weights (a second SSD) |
OLLAMA_HOST |
Address to listen on (for example 0.0.0.0 for the LAN) |
A one-off test in a PowerShell session, before you make it permanent:
$env:OLLAMA_MODELS="D:\ollama-models"
Create D:\ollama-models, quit and restart the Ollama service (taskbar icon), then test:
ollama pull llama3.2:3b-instruct-q4_K_M
Save it for your user (after the test works):
[System.Environment]::SetEnvironmentVariable('OLLAMA_MODELS', 'D:\ollama-models', 'User')
Close PowerShell, restart the service, then:
ollama list
Set OLLAMA_HOST=0.0.0.0 only on a home network you trust, with the firewall on, and without casually forwarding port 11434 on the router. For a container stack instead of the native Windows service, see Docker Ollama + Open WebUI.
What to do next
Once the CLI checks out, continue in this order, or jump to the job you actually want:
- Tags and quantization — Choose a first model: keeps a heavy 13B off a 16 GB machine, and separates instruct from base.
- A family interface — Open WebUI for beginners: the same Ollama engine, chat in the browser on the home LAN.
- A graphical app with no server — Install LM Studio: you can compare side by side if you watch disk space (SSD for models).
- Code and scripts — A local coding assistant with Continue once VS Code is part of how you work.
- Another operating system — Ollama on Linux and Mac for a second machine in the house.
- House rules — A private family assistant: who can reach what, with port 11434 kept off the public internet (see also privacy when AI stays home).
- The full path — Getting started hub: beginner and intermediate tables, and durations, so you do not reread the same links scattered through this tutorial.
Hardware (only if the test fails)
The useful order: measure on the PC you have first. If steps 1 to 4 succeed, you do not need to buy anything. If a short reply takes more than a minute on CPU alone, read VRAM and models to find the bottleneck (RAM, VRAM, disk), then the hardware selector: usage questions and budget before any order. The Amazon link below opens a sample search (mini PCs, October 2026). Compare the spec sheets yourself. Mini PC search on Amazon (Amazon affiliate link: a commission is possible at no extra cost to you. Check the current price on Amazon — how affiliation works).
Power use and noise
A laptop will spin its fans hard under CPU load. A quiet mini PC can be worth it if Ollama runs around the clock. Estimate the kWh with the electricity budget and the 3-year cost comparison.
Safety
Ollama listens locally (port 11434). Leave that port closed on the router until the service has a login. For access away from home, a household VPN is the calmer choice, ahead of opening the port raw. Prompts stay on your disk: ordinary backups and Windows updates still apply. If you are unsure about the data you would process (an employer’s code, health records), reread privacy and local AI before you index a whole repository in Continue.
Answers to the Windows questions (WSL2, the .ollama folder, uninstalling, a failed pull) are in the FAQ at the bottom of the page.
Common questions
Do I need WSL2 for Ollama on Windows?
No, not for the recent native Windows installer. Ollama runs as a Windows service. WSL2 is still useful for other Docker tools. It is not required for Ollama by itself.
Where are the models stored?
In your user profile, in the .ollama folder (the models are large). Plan on 10–50 GB on an SSD.
How do I uninstall it cleanly?
Settings, Apps, Ollama, Uninstall. Then delete the .ollama folder if you want the disk space back.
Why does ollama pull fail?
Check the internet connection, Windows Firewall, a company proxy, and free disk space on the drive that holds your user profile. If the message names an unknown model, copy the tag from the text block in step 3 (for example llama3.2:3b-instruct-q4_K_M, with no spaces). See choose a first model.
Sources
Narrow it down with the hardware selector · French version of this section