How to Install Ollama on Windows 11 & Windows 10 (WSL2 & Native Setup)

Ollama is the easiest way to run open-weight models on Windows, and the core install really is one file. What trips people up is everything after the installer: whether the GPU is actually being used, whether your PATH picked up the new command, where the models land on your C: drive, and whether you should run it inside WSL2 instead. This guide assumes you want more than a toy and walks the native path first, then the WSL2 path, with the checks that tell you it’s working. For the broader architecture and command reference, start with the Ollama Complete Guide.

1. What you need before you install

Ollama runs without a GPU, but if you have a capable NVIDIA card you should size things realistically rather than install first and hope. The pattern for Windows is the same as everywhere: weights + ~2 GB context cache + a little headroom is what VRAM you need. That means:

If you already know your hardware, skip ahead. If you’re unsure whether a model fits, an 8B Q4 model needs about 6–7 GB of VRAM — that’s the entry point to “a real coding/chat model” on a consumer card.

2. Native install (recommended for most people)

Install

1. Download OllamaSetup.exe from ollama.com (it’s the same installer for Windows 10 and 11).
2. Run it and wait — there’s no options screen; it installs the binary and registers a background service/tray app automatically.

Verify the command is on PATH

Open a new PowerShell or Windows Terminal window (the PATH only refreshes in new sessions) and check:

ollama --version

If you get ollama version is 0.5.x (or newer), you’re set. If you get “not recognized,” see Section 6 — the most common cause is an un-refreshed terminal, not a broken install.

Pull and run your first model

ollama run llama3.2

It downloads the 3B weights (about 2 GB) with a progress bar, then drops you into an interactive >>> chat prompt. Type a question, and you will see output stream. Type /bye or press Ctrl+D to exit.

3. Confirm the GPU is actually doing the work

The biggest silent failure on Windows: Ollama falls back to CPU, and nobody notices until generation is painfully slow. While a model is loaded, open a second terminal and run:

ollama ps

This shows what’s in memory and which processor runs it:

NAME               ID              SIZE      PROCESSOR    UNTIL
llama3.2:latest    a80c4f172d6e    2.0 GB    100% GPU     4m52s

• 100% GPU — great, it’s using your NVIDIA card (or the iGPU via the fallback on cards without CUDA).
• 100% CPU — the GPU isn’t being used. Fix in Section 5.
• a split like 52%/48% GPU/CPU — the model plus context overflowed VRAM; either drop context or use a smaller model.

Note that on Windows, default-built Ollama uses CUDA for NVIDIA and falls back to a CPU build otherwise. On a dual-GPU laptop, “the GPU” it picks may be your weak integrated one if the big NVIDIA GPU isn’t selected in Windows graphics settings — a genuinely common culprit for “why is it slow on my gaming laptop?”

4. Changing where Windows stores models

By default models go to your user profile, which on most machines means the C: drive:

C:\Users\<YourUsername>\.ollama\models

Models run from 4 GB to 40+ GB, so three or four of them can fill a small boot SSD. To relocate:

1. Finish running any model, then quit Ollama (right-click the tray llama icon → Quit).
2. Open System settings → search “environment variables” → Edit environment variables for your account.
3. Add a user variable OLLAMA_MODELS pointing at e.g. D:\OllamaModels (create that folder).
4. Optionally copy the existing .ollama\models contents there so you don’t re-download.
5. Restart Ollama and run ollama list to confirm your models still appear.

A full walkthrough of moving and managing the model store, including the caveats, is in the Ollama Model Storage guide.

5. GPU problems: why it’s on CPU, and how to fix it

If ollama ps shows CPU when you have an NVIDIA GPU, work through these in order:

• Driver too old. Run nvidia-smi — the “CUDA Version” at the top of its output reflects your driver’s CUDA capability. Update to a recent Game Ready or Studio driver (550+).
• Laptop picked the wrong GPU. Settings → System → Display → Graphics, choose the Ollama/your app, and force “High performance” (the dGPU).
• Another process holds VRAM. Browser tabs and games can eat enough VRAM that Ollama decides it can’t fit and backs off. Close heavy apps, then ollama stop llama3.2 and re-run.

If the model itself is bigger than your card, that’s not a bug — drop to a smaller quant or a smaller model. The GPU-and-VRAM section of the complete guide has the sizing math.

6. Common Windows snags and fixes

For connection-refused, port conflicts, and deeper diagnostics, the Ollama troubleshooting guide is the full manual.

7. The WSL2 route (and when to use it)

Install Ollama inside WSL2 (an Ubuntu environment on Windows) when you want Linux-native tooling: Docker containers, systemd services, or a headless server setup that mirrors production. The native Windows app is a user process; WSL2 gives you the same daemon-as-service experience you’d get on a Linux box.

Check GPU passthrough first

Modern WSL2 passes your NVIDIA GPU through automatically if you have the driver installed on Windows. Verify from inside the WSL terminal that the GPU is visible before installing Ollama:

nvidia-smi

If nvidia-smi runs and lists your card, passthrough works. Then install normally:

curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.2

Two practical notes on the WSL2 path. First, the Linux service stores models under /usr/share/ollama/.ollama/models (root-owned), not your Windows profile — plan disk accordingly. Second, WSL2 has its own network namespace: the Windows-native daemon binds 127.0.0.1 in Windows, while a WSL2 daemon binds inside WSL. They don’t automatically share a port, so pick one environment for serving and point clients at it consistently.

Frequently asked questions

Does Ollama start automatically on Windows?

Yes — the installer adds it to Windows startup, so the local API is already listening on http://127.0.0.1:11434 after you log in. Disable it under Task Manager → Startup Apps if you’d rather launch it manually.

Can I use an AMD GPU on Windows?

There’s preview support for select AMD Radeon cards, but on Windows the most stable path is running Ollama in WSL2/Linux where ROCm is better supported. For a reliable AMD experience, go the WSL2 route in Section 7.

How do I connect a UI like Open-WebUI or AnythingLLM?

Point the app at Ollama’s endpoint http://127.0.0.1:11434. The full setup is in Connecting AnythingLLM and Open-WebUI to Ollama.