Ollama runs on Windows natively: download OllamaSetup.exe, run it, then ollama run <model>.
No administrator rights, no WSL, nothing leaves the machine.
1. Check what Windows needs
The floors come from Ollama’s Windows documentation:
| Requirement | What the docs state |
|---|---|
| Windows | 10 22H2 or newer, Home or Pro |
| NVIDIA driver | 551.61 or newer |
| AMD | a ROCm v7 / HIP7-capable or Vulkan-capable driver stack |
| Free disk | at least 4 GB for the binaries, plus room for models |
2. Install Ollama
irm https://ollama.com/install.ps1 | iex
What it does: the one-liner from the download page; OllamaSetup.exe is the same install with a
wizard. It lands in your home directory, so no elevation prompt.
Short on space on C:? The installer takes a directory:
.\OllamaSetup.exe /DIR="D:\ollama"
What it does: puts the binaries where you point it.
3. Pull and run your first model
ollama run gemma4
What it does: downloads the model on first use — several GB — then opens a chat; /bye exits.
Any library tag works in its place; check its size against
what fits in your RAM or VRAM and
the quantization you pick.
4. Verify it worked
ollama ps
What it does: lists the models loaded right now, with the memory each holds. Run it while a reply is still streaming — an idle Ollama has nothing loaded.
curl.exe http://localhost:11434/api/tags
What it does: asks the local server for its models. JSON back means Ollama
serves on port 11434 and your code can call that API. Use curl.exe,
not curl — PowerShell aliases the bare name to Invoke-WebRequest. Git Bash has the real one.
5. Move the model store off C:
Models land in %HOMEPATH%\.ollama and grow fast. The docs move the store with the
OLLAMA_MODELS variable, set through Windows Settings — the same User variable from a terminal:
[Environment]::SetEnvironmentVariable('OLLAMA_MODELS', 'D:\ollama-models', 'User')
What it does: points the store at another drive. Quit Ollama from the tray and start it again; the server reads the variable at startup.
Failure mode: it answers, but at one word a second
That is the CPU doing the work. Two causes worth checking: a graphics driver below the floors above, and a model bigger than your VRAM. Update the driver first; if nothing changes, drop to a smaller tag or a lower quantization.
Ollama’s Windows page documents the native build only. The Linux build inside WSL 2 is a separate install with its own model directory.