How to Install Ollama on Windows

If you have read our broader overview of how to run local AI models (LLMs) on Windows, Ollama was one of the two tools we recommended. Here we go deeper on Ollama specifically. It is the favourite for people who are comfortable with a command line, or who want to wire a local model into other software.

What is Ollama?

Ollama is an open-source runner for large language models. Think of it as a manager and engine rolled into one: it downloads models for you, keeps them organised on disk, loads them into memory when you want to chat, and exposes them through both a simple command line and a local web API. It handles the awkward parts — detecting your hardware, choosing how much of a model to put on the GPU, managing the model files — so you do not have to.

Crucially, everything runs on your machine. After a model file is downloaded, no internet connection is needed and nothing you type is sent anywhere. That makes Ollama a strong choice for confidential documents, offline work, and anyone who would rather not send their prompts to someone else's server. It installs as a native Windows program with no Python environment or extra dependencies to wrangle.

Download and install on Windows

Installing is deliberately painless — there is no build step and nothing to configure.

1. Go to the official Ollama website and open its download page. Choose the Windows download; you will get an installer named OllamaSetup.exe. Only ever download it from the official site, not a random mirror.
2. Double-click the installer and follow the short wizard. It installs the Ollama program and starts its background service automatically.
3. When it finishes, Ollama runs quietly in the background — you will usually see its icon in the system tray. There is no window to open yet; you drive it from a terminal.

That is the whole installation. If you prefer the command line even for setup, Ollama also publishes a PowerShell one-liner on its download page that fetches and runs the same installer, but for most people the double-click .exe is simplest.

Run your first model

Now open a terminal. On Windows 11, Windows Terminal or PowerShell both work — press the Start button, type "terminal" or "powershell", and open it. Then run a model with a single command:

ollama run llama3.1

The first time you run a given model, Ollama downloads it — you will see a progress bar as it pulls the file. Once the download finishes it drops you straight into an interactive chat prompt. Type a question, press Enter, and the model answers locally. To leave the chat, type /bye or press Ctrl+D.

Run the same command again later and it skips the download, because the model is already on disk — it just loads and starts chatting. A small model in the 7–8B range is a sensible first choice: it downloads quickly, fits on modest hardware, and is capable enough for everyday drafting, summarising and coding help.

Pulling and managing models

Over time you will collect several models and want to keep them tidy. A handful of commands cover almost everything:

ollama pull mistral      # download a model without chatting
ollama list              # show every model you have on disk, with sizes
ollama ps                # show which models are currently loaded in memory
ollama rm llama3.1       # delete a model and reclaim its disk space
ollama show llama3.1     # print details about a model

Because models are large, ollama list and ollama rm are the two you will reach for most — the first to see what is eating your disk, the second to clear out anything you no longer use. By default the model files live inside your user profile on the system drive, under %USERPROFILE%\.ollama\models. If that drive is tight, you can point Ollama at a different folder by setting the OLLAMA_MODELS environment variable to a path on another drive — both the default Windows path and that variable are documented in the Ollama FAQ.

The model library

Ollama maintains a curated library of models you can pull by name, published on its website. It spans many of the well-known open families — general chat models, smaller lightweight models for weaker hardware, larger models for stronger machines, and specialised ones tuned for coding or other tasks. Browsing the library on the site shows you each model's available sizes and quantizations, and gives you the exact name to type after ollama run or ollama pull.

Two ideas are worth understanding when you pick a model. Parameter count — the number before the "B", as in 7B or 70B — is a rough proxy for capability and for how much memory a model needs; bigger is generally smarter but heavier. Quantization is a compressed version that stores the model's numbers at lower precision, shrinking the file and the memory it needs for a small drop in quality. A 4-bit quantized model is the usual sweet spot, and Ollama defaults to sensible quantized versions so you rarely have to think about it. For a fuller explanation of matching a model to your RAM and VRAM, see our guide to running local LLMs on Windows.

GPU acceleration

Ollama runs on the CPU by default if it finds no supported graphics card, using ordinary system RAM. That works, but a GPU is dramatically faster because it can hold the model in its own high-speed memory and do the maths in parallel. The good news is you do not have to configure anything: Ollama detects a supported GPU and uses it automatically.

• NVIDIA is the best-supported and most trouble-free path on Windows. Ollama accelerates modern NVIDIA GPUs through CUDA; just keep your Game Ready or Studio graphics drivers current.
• AMD Radeon RX and PRO discrete cards are supported through ROCm on Windows, with the newer RDNA generations best covered. Make sure you have up-to-date AMD drivers.
• If a model is larger than your available VRAM, Ollama offloads the overflow to system RAM. It still runs, just more slowly, since part of the work falls back to the CPU.

The single biggest factor in how large a model you can run smoothly is how much VRAM your GPU has. As a rough guide, an 8 GB card comfortably handles small 7–8B models, a 12–16 GB card steps up to mid-size models, and 24 GB or more opens the door to larger ones. You can confirm Ollama is actually using the GPU by running a model and checking your GPU's utilisation in Task Manager while it replies.

The built-in API

One of Ollama's most useful features is invisible at first: while it runs, it serves a local HTTP API on your own machine at http://localhost:11434. That is what lets other software use your local models without any of the conversation leaving the PC. Chat front-ends, note-taking apps, coding assistants and your own scripts can all point at that address.

The main endpoints are /api/generate (a single completion for a prompt) and /api/chat (a back-and-forth conversation), with others for pulling, listing and managing models. A quick way to prove it is working, from PowerShell:

curl http://localhost:11434/api/tags

That returns the list of models you have installed, in JSON. Because the server listens only on your own machine by default, it is not exposed to the internet unless you deliberately change that — the Ollama FAQ confirms it binds 127.0.0.1 port 11434 by default and that you change this with the OLLAMA_HOST variable — so your local setup stays private out of the box. Ollama's API also has an OpenAI-compatible mode, which means many tools written for OpenAI's API can be pointed at your local Ollama with only a change of address.

Adding a graphical interface

Ollama is command-line first, but you do not have to live in a terminal. Recent versions of Ollama ship with an official desktop app for Windows that gives you a simple chat window on top of the same engine, so you can type to a model without touching PowerShell.

Beyond that, because Ollama exposes the local API described above, a whole ecosystem of third-party interfaces can connect to it — browser-based chat front-ends and desktop clients that give you conversation history, model switching and settings in a friendly window while Ollama does the work underneath. If you would rather have a graphical app that bundles model browsing and chat together from the start, LM Studio is the popular alternative; we compare the two in detail in Ollama vs LM Studio.

Uninstalling

Removing Ollama is as ordinary as removing any other Windows program:

1. Open Settings and go to Apps › Installed apps.
2. Find Ollama in the list, click the three-dot menu (or the entry) and choose Uninstall.
3. That removes the program itself. Your downloaded models may be left behind, because they live separately and can be large. To reclaim that space, delete the Ollama models folder in your user profile after uninstalling.

That is everything. From here, the natural next steps are to try a couple of different models to see what your hardware handles comfortably, and to connect a friendlier chat interface if you prefer clicking to typing. For the wider picture of local AI on Windows — including Microsoft's own on-device stack — see our overview of Windows AI and the guide to running local LLMs on Windows. If you are curious about the small, efficient models designed for on-device use, our explainer on what Liquid AI is is a good companion read.