Install Ollama on Windows

What you will build

By the end of this tutorial, your Windows PC should have:

• Ollama installed as a native Windows application;
• the ollama command available from PowerShell or Command Prompt;
• one downloaded model that can answer prompts locally;
• a local Ollama API available at http://localhost:11434;
• a repeatable check for installed and currently running models.

The important part is not merely completing the installer. We will verify each layer separately: application → CLI → model → API.

Requirements

Ollama's current Windows documentation lists Windows 10 22H2 or newer, Home or Pro. Ollama runs as a native Windows application and supports hardware acceleration on supported NVIDIA and AMD Radeon configurations.

The installer itself needs at least 4 GB of disk space, according to the official documentation. Model storage is additional and can grow to tens or hundreds of gigabytes depending on what you download.

You do not need Administrator rights for the normal OllamaSetup.exe installation; it installs in the user account by default.

1. Install Ollama for Windows

Use the official Windows installer:

Open official Windows download

Run OllamaSetup.exe and complete the installation.

The normal installer does not require Administrator privileges. After installation, Ollama runs in the background and adds its command-line program to your user PATH.

Checkpoint

Open a new PowerShell window after installation. Starting a fresh terminal matters because an already-open terminal may not see a newly updated PATH.

2. Verify the Ollama CLI

Run:

ollama --help

If the command is available, the terminal should print Ollama's command help rather than a “command not found” / “not recognized” error.

Now list locally installed models:

ollama ls

On a fresh installation the list may be empty. That is fine: this checkpoint tests the CLI, not the model.

3. Download and run a model

Ollama's CLI reference uses ollama run <model> to run a model. Choose a model that is appropriate for your hardware from Ollama's model library rather than automatically choosing the largest one.

The general command is:

ollama run <model>

For example, after selecting a suitable model name from the current library:

ollama run MODEL_NAME

On first use, Ollama can download the required model data. Once the interactive prompt appears, ask a short deterministic test question such as:

Reply with exactly: LOCAL AI WORKS

Do not treat exact wording from a generative model as a system-health requirement. The real checkpoint is that inference completes and returns a response.

To download without immediately starting an interactive session, the CLI also provides:

ollama pull <model>

4. Verify the local HTTP API

On Windows, the local Ollama API is served at:

http://localhost:11434

The Ollama API base path is:

http://localhost:11434/api

A convenient first test does not require generating another response. Ask the API for the local model list:

Invoke-RestMethod -Uri http://localhost:11434/api/tags

If Ollama is running, PowerShell should receive structured data from the local service.

Now test generation using the model you downloaded:

$body = @{
  model  = "MODEL_NAME"
  prompt = "Reply with exactly: API WORKS"
  stream = $false
} | ConvertTo-Json
Invoke-RestMethod   -Method Post
-Uri http://localhost:11434/api/generate   -ContentType "application/json"
-Body $body

Replace MODEL_NAME with the installed model name.

For local inference, Ollama's API documentation uses the local server and does not require the cloud API-key authorization header.

5. Inspect installed and running models

List downloaded models:

ollama ls

List models currently loaded/running:

ollama ps

Stop a running model when needed:

ollama stop <model>

These commands are useful later when comparing memory use because “downloaded”, “running” and “loaded into GPU memory” are different states.

Changing where models are stored

By default, downloaded models can consume a lot of space in the user environment. Ollama documents the OLLAMA_MODELS user environment variable for changing the model-storage location on Windows.

In Windows Settings, search for environment variables, edit the variables for your account, and create or change:

OLLAMA_MODELS=D:\AI\OllamaModels

Use a path that actually exists on your machine. If Ollama is already running, quit the tray application and relaunch it after changing the variable.

Troubleshooting

ollama is not recognized

Close the current terminal and open a new PowerShell window. The installer places the Ollama binaries under the user's local application directory and adds them to the user PATH.

The official documentation identifies the binary location as:

%LOCALAPPDATA%\Programs\Ollama

The API does not answer on port 11434

First confirm that Ollama is running in the Windows tray. Then try:

Invoke-RestMethod -Uri http://localhost:11434/api/tags

For troubleshooting, Ollama documents its Windows logs under:

%LOCALAPPDATA%\Ollama

server.log contains the most recent server logs and app.log contains GUI application logs.

The model is too slow or runs out of memory

Do not immediately assume the installation failed. Model size, quantization, context length, GPU support and CPU/GPU placement all affect performance and memory use.

Check:

ollama ps

Then compare the model with the E—DOCEO VRAM and Quantization guides. A smaller or differently quantized model may be a better diagnostic test.

You see square characters in the Windows terminal

Ollama's Windows documentation notes that Unicode progress characters can appear as unknown squares with some older Windows 10 terminal fonts. Change the terminal font before treating this as an Ollama failure.

E—DOCEO verification checklist

The September 18, 2026 test completed the core verification path used by this tutorial:

• [x] Windows test environment recorded.
• [x] Windows version, CPU, RAM and GPU recorded.
• [x] Ollama version recorded: 0.34.2.
• [x] CLI available in PowerShell.
• [x] ollama ls and ollama ps completed.
• [x] Exact model recorded: gemma3:1b, ID 8648f39daa8f.
• [x] First local inference completed.
• [x] /api/tags responded locally.
• [x] /api/generate completed with stream: false.
• [x] Runtime placement recorded as 100% GPU, context 4096.
• [x] GPU-memory snapshots recorded before and after model loading/inference.
• [x] Three API generation runs recorded for the Lab benchmark.

The detailed measurements, methodology and limitations are published in Gemma 3 1B on GTX 1060 3GB: E—DOCEO Lab Test.

Continue learning

Read What Is Ollama? for the architecture behind this setup, then continue to VRAM and Quantization before experimenting with substantially larger models.

Primary sources

• Ollama — Windows
• Ollama — CLI Reference
• Ollama — API Introduction
• Ollama — API model list
• Ollama — Hardware support
• Ollama — Context length