How to Install Ollama on Windows

Download the Windows installer, confirm PATH is set, verify the local API server is up, and understand when WSL2 gives you better GPU performance.

10 minBeginner

Ollama has shipped a native Windows installer since version 0.1.9. It installs an ollamad system service, sets up PATH, and bundles the llama.cpp backend with CUDA support for NVIDIA GPUs. You do not need WSL to get started, though WSL2 is worth considering if you run into GPU sharing issues.

What you need

Windows 10 22H2 or Windows 11 (64-bit)
At least 8 GB RAM (16 GB recommended for 8B models)
NVIDIA GPU with CUDA 11.8+ drivers for GPU acceleration (CPU-only also works)
About 5 GB of free disk space per model
No administrator rights required for the user install

Step 1: Download and run the installer

Go to ollama.com and click the Download for Windows button. This gives you OllamaSetup.exe, a standard Windows installer under 500 MB. Run it and accept the UAC prompt. The installer does three things: extracts the binary to %LOCALAPPDATA%\Programs\Ollama, adds that directory to your user PATH, and registers a Windows service that starts on login.

PowerShell - verify install

$ollama --version

ollama version 0.9.x

$$env:PATH -split ';' | Select-String 'ollama'

C:\Users\you\AppData\Local\Programs\Ollama

If ollama is not found after install

Close and reopen PowerShell or Command Prompt. The installer modifies the user PATH but the change only takes effect in new shell sessions. If it is still missing, open System Properties > Environment Variables and confirm %LOCALAPPDATA%\Programs\Ollama appears in the user PATH entry.

Step 2: Pull and run a model

Open a new PowerShell or Command Prompt window. The ollama run command downloads the model on first use and then opens an interactive prompt. Use llama3.3 for a capable general-purpose model, or phi4-mini for something that runs faster on lower-end hardware.

PowerShell - first model

$ollama run llama3.3

pulling manifest...

pulling model weights... ████████████ 100% 4.7 GB

>>> Send a message (/? for help)

$Summarize what a transformer architecture does.

A transformer is a neural network that...

Type /bye to exit, or press Ctrl+D

Step 3: Verify the local server

Ollama starts an HTTP API on localhost:11434. Check it responds before pointing any client library at it. PowerShell's Invoke-WebRequest works, or use curl.exe which ships with Windows 10 1803 and newer.

PowerShell - verify API

$curl.exe http://localhost:11434

Ollama is running

$curl.exe http://localhost:11434/api/tags

{"models":[{"name":"llama3.3:latest","size":4661211648}]}

Step 4: NVIDIA GPU check

If you have an NVIDIA GPU, Ollama automatically uses CUDA when the driver version is 525 or higher. Confirm it is being used by checking the logs or running a quick generation and watching GPU load in Task Manager.

PowerShell - GPU check

$nvidia-smi

NVIDIA-SMI 555.xx Driver Version: 555.xx CUDA Version: 12.5

During model inference, GPU Memory-Usage should spike above zero

$nvidia-smi dmon -s u

# gpu sm mem enc dec

# 0 62 85 0 0 (active during inference)

When to use WSL2 instead

The native Windows install works well for most use cases. Switch to WSL2 if you need to share the GPU between a Docker container and Ollama simultaneously, or if you are already running your development environment inside WSL2 and want everything in one place. In WSL2, install Ollama with the Linux shell script: curl -fsSL https://ollama.com/install.sh | sh

Useful management commands

ollama list — show downloaded models
ollama rm modelname — free disk space by removing a model
ollama ps — show currently loaded models and their memory usage
ollama serve — start the server manually if the service is not running
Get-Service ollamad — check the Windows service status in PowerShell

Ollama is now running as a Windows service and the API is live on port 11434. From here you can pull any model from the Ollama model library at ollama.com/library and use it from Python, Node, or any OpenAI-compatible SDK without sending data to an external server.

Local models run better with more VRAM. CompareRTX GPUs on Amazonbefore you upgrade.(affiliate link. We may earn a commission at no extra cost. Disclosure)