Beginner 11 minOllama

Install Ollama in 5 minutes (Windows, macOS, Linux)

Direct response

To install Ollama, download the installer from ollama.com (Windows, macOS 14 or later) or run the official curl command on Linux, then type ollama run followed by a model name. Allow 5 minutes for the hands-on steps, excluding the model download. No GPU is required: available memory determines the possible model size.

This page covers installing Ollama on all three systems, what to check before you begin, the first model to run, and a few settings that prevent unpleasant surprises: where models are stored, how to free memory, and what remains local. System-specific details are covered in a dedicated guide.

By Mohamed Meguedmi·Update 2026-09-29·Tested on Windows 11

#Install Ollama: the short version for every system

Ollama is an engine that downloads open-weight language models and runs them on your machine, with a command-line interface and a local API. Installation takes a single step on each system. On Windows and macOS, you download an application from ollama.com/download; on Linux, an official script installs the program and service. Then the ollama run suivie command with a model name downloads it on the first call and opens a conversation.

Installing Ollama by operating system (official documentation, September 2026)
SystemMethodRequirements announced by Ollama
WindowsOllamaSetup.exe installer, or PowerShell command irm https://ollama.com/install.ps1 | iexWindows 10 22H2 or later; driver NVIDIA 551.61 or later if NVIDIA card
macOSollama.dmg file dragged into Applications, or brew install ollamamacOS 14 Sonoma or later; Apple M chip (CPU and GPU) or x86 (CPU only)
Linuxcurl -fsSL https://ollama.com/install.sh | shDriver NVIDIA 550 or later (or ROCm v7 for AMD) if GPU acceleration is desired

In all three cases, the server listens on http://localhost:11434, and the ollama command is available in the terminal. The following sections cover each system without repeating the dedicated step-by-step guides, then move on to the first model.

#Prerequisites: which machine, which model size

The Local AI Kit

Your private ChatGPT, free, on your own machine in an hour — LM Studio, Ollama, Open WebUI, your documents, no cloud.

  • Lifetime online access
  • PDF + files
  • Lifetime updates
The Local AI Kit →or all kits for life — $49 →

Ollama runs on almost any recent machine; what varies is the acceptable model size. The limiting factor is the memory available for the weights: the VRAM of a dedicated graphics card, the unified memory of a Mac Apple Silicon, or, failing that, the system RAM, in which case only the CPU is used and responses are slower.

Memory reference by model size, in Q4 quantization (weights only)
Model sizeMemory for the weightsReasonable starting machine
3Babout 2 GBPC with 8 GB of RAM, no GPU
7-8Babout 5 GB16 GB of RAM, or 8 GB GPU
14Babout 9 GB12 GB GPU, or 16 GB Mac
32Babout 19–20 GB24 GB GPU, or 32 GB Mac
70Babout 40 GB64 GB Mac or more, or multiple GPUs

These figures count only the weights. You also need to add the context (KV cache), which grows with conversation length, and leave room for the system: a 5 GB model on an 8 GB machine works, but leaves no headroom for a browser open alongside it. For choosing a quantization, see the guide to Q4, Q5, and Q8.

For disk space, the Ollama documentation states that the program requires at least 4 GB on Windows, plus the models, which can reach tens or even hundreds of GB depending on what you download. An internal SSD is preferable: loading a model from a slow external drive is noticeable every time you start up.

→
No GPU: it's possible, but choose a small model
On the CPU alone, a 3- to 4-billion-parameter model remains usable for summarizing or writing a short text. Beyond 7-8B without an accelerator, the wait becomes painful. The guide to LLMs without a GPU breaks down models by RAM capacity.

#Install Ollama on Windows

Ollama works as a native Windows application, supporting NVIDIA and AMD Radeon cards. Once installed, it runs in the background, and the ollama command becomes available in cmd, PowerShell, or any other terminal. Installation does not require administrator privileges and defaults to your personal folder.

  1. 01
    Download
    On ollama.com/download, choose Windows and download OllamaSetup.exe. The page also offers the PowerShell command irm https://ollama.com/install.ps1 | iex, which does the same thing.
  2. 02
    Launch the installer
    Double-click the file. To install it somewhere other than the user directory, run it with the /DIR parameter followed by the path, for example OllamaSetup.exe /DIR="d:\dossier".
  3. 03
    Check
    Open PowerShell and type ollama --version. A Ollama icon also appears in the notification area, near the clock.
  4. 04
    Check the GPU driver
    With a NVIDIA card, the documentation requires driver 551.61 or newer. For an AMD card, you need either the ROCm v7 stack or a Radeon driver compatible with Vulkan.
i
Windows 10: terminal and fonts
The documentation notes that Unicode characters in the progress bar may appear as squares in some Windows 10 terminals. Changing the terminal font fixes the problem; it’s cosmetic and has no effect on the download.

For the complete step-by-step guide and Windows-specific troubleshooting, see the guide dedicated to Windows 11.

#Install Ollama on macOS

Ollama requires macOS 14 Sonoma or later. Apple M chips benefit from GPU compute; Intel (x86) Macs run only on the CPU, limiting them to very small models. The method recommended by the documentation is to mount the ollama.dmg file and drag the application into the system’s Applications folder.

Official application
On first launch, the application checks whether the ollama command is in your PATH and, if not, asks for permission to create a link in /usr/local/bin.
Homebrew
An ollama formula is available in Homebrew (brew install ollama): it’s the terminal option for anyone who already has Homebrew, without using the .dmg file.
Unified memory
On Apple Silicon, the CPU and GPU share the same memory: a 16 GB Mac can therefore load a 14B model in Q4, whereas a PC with an 8 GB VRAM GPU could not.

The dedicated macOS guide explains how to choose a model based on your Mac's memory and the useful settings on Apple Silicon.

#Install Ollama on Linux

On Linux, one command is enough. The official script downloads the program and configures it as a service.

Official installation
curl -fsSL https://ollama.com/install.sh | sh

The documentation also offers a manual installation: extract the ollama-linux-amd64.tar.zst archive into /usr (or the arm64 archive for ARM machines), add the extra ROCm archive for AMD cards, then create a user and a systemd service if you want automatic startup. To update, simply run the installation script again; the OLLAMA_VERSION variable lets you install a specific version.

For a NVIDIA card, install the driver and check with nvidia-smi, which should display your card. For an AMD card, the documentation requires the ROCm v7 driver. The dedicated Linux guide covers systemd, permissions, and environment variables.

#Run your first model

Once Ollama is installed, ollama run t downloads the requested model if it is missing, then opens a conversation in the terminal. The first call depends on your connection: a 3 to 5 GB model takes a few minutes, and the files then remain on disk. Subsequent calls start without another download; only loading it into memory remains.

First launch
ollama run qwen3.5:4b

The Qwen 3.5 family is published on ollama.com in sizes from 0.8b to 122b, with vision, tools, and reasoning depending on the version. The 4b tag fits in a few GB in Q4, so it works on almost any machine. If you have 16 GB of memory, try an 8–9B model; the table above provides the reference point. The site's VRAM-based selection pages can help you narrow it down.

→
Leave the conversation
Type /bye to exit. By default, Ollama keeps the model in memory for 5 minutes after the last request, which speeds up subsequent launches; ollama stop followed by the model name unloads it immediately.

#Commands to know after installation

ollama list
Displays downloaded models and their size on disk.
ollama pull nom
Downloads a model without running it, which is useful for preparing several models before a trip.
ollama ps
Shows the models loaded in memory and whether each one runs on the GPU or CPU.
ollama stop nom
Unloads a model from memory without waiting the default 5 minutes.
ollama rm nom
Delete a model to reclaim disk space.
ollama serve
Starts the server manually, useful on Linux if the service is not running.

For what comes next (updating, completely removing Ollama and its models), see the guide to maintaining Ollama.

#Where models are stored and how to move them

Models quickly weigh several dozen GB, and the system drive is rarely the best place for them. The official FAQ gives their default locations: ~/.ollama/models on macOS, /usr/share/ollama/.ollama/models on Linux, C:\Users\%username%\.ollama\models on Windows.

To move them, set the OLLAMA_MODELS environment variable to the desired folder. On Windows, quit Ollama from the notification area, open your account’s environment variables, create the variable, then relaunch the application. On Linux with the standard installation, the ollama user must have read and write permissions for this folder: the documentation suggests sudo chown -R ollama:ollama followed by the folder.

#Privacy: what stays local

In local mode, prompts and responses stay on your machine: Ollama's FAQ states that the company sees neither your prompts nor your data when you run a model locally. Cloud models are different: they go through Ollama's servers. If you want strictly local use, the documentation describes a local-only mode: the OLLAMA_NO_CLOUD=1 variable, or the disable_ollama_cloud setting in ~/.ollama/server.json. Then restart Ollama.

By default, the server listens only on 127.0.0.1, port 11434: no one else on your network can query it. Changing OLLAMA_HOST to 0.0.0.0 exposes it to the entire local network, without authentication. Do this only behind a firewall or proxy.

#Troubleshooting: the five common problems

Command not found
Close and reopen the terminal: the installer just modified PATH. On macOS, restart the application so it offers the /usr/local/bin link.
GPU not used
Launch a model, then run ollama ps: the processor column indicates GPU or CPU. Update the driver (NVIDIA 551.61 minimum on Windows, 550 on Linux), then relaunch Ollama.
AMD card not recognized on Windows
The documentation specifies that certain RDNA2 systems (RX 6000) do not expose ROCm v7 under Windows. Vulkan, enabled by default, is therefore the recommended fallback.
Port 11434 is occupied
Another instance of Ollama is already running, or another program is using the port. Quit the application, then restart it.
Download blocked behind a proxy
Use HTTPS_PROXY and avoid HTTP_PROXY, which can disrupt client connections according to the FAQ; the proxy certificate must be installed as a system certificate.
Frequently asked questions about installing Ollama
How do you install Ollama on Windows 10?+
Ollama requires Windows 10 22H2 or later, in the Home or Pro edition. Download OllamaSetup.exe from ollama.com/download and run it: administrator privileges are not required. Then open PowerShell and type ollama --version. With a NVIDIA card, install driver 551.61 or later first; otherwise Ollama falls back to the processor and responses slow down significantly.
Is Ollama free?+
Yes for local use: the program and the open models it runs are free; only the electricity used by your hardware costs money. Ollama also offers cloud models with an account, separate from running models on your machine; they are part of a separate offering and aren't necessary for this guide.
Do you need a GPU to use Ollama?+
No. Ollama runs on the CPU alone, but responses are slower, limiting comfortable use to small models with 3 to 4 billion parameters. An NVIDIA or compatible AMD card, or a Mac with Apple Silicon, significantly speeds up generation. Then check with ollama ps where the model is actually loaded.
Where does Ollama store models?+
On macOS, ~/.ollama/models; on Linux, /usr/share/ollama/.ollama/models; on Windows, C:\Users\nom\.ollama\models. To change the location, set the OLLAMA_MODELS variable, then restart Ollama for the change to take effect. Next, copy the existing models to this folder, or redownload them if you prefer to start from scratch, then verify with ollama list.
Does Ollama send my conversations over the Internet?+
Not in local mode: Ollama states that it does not see your prompts or data when the model runs on your machine. Cloud models, however, pass through its servers. OLLAMA_NO_CLOUD=1 disables these features to remain strictly local; then restart Ollama to apply the setting.
How do you know whether Ollama is using my GPU?+
Start a model, then type ollama ps in a second terminal. The processor column shows the share of the model loaded on the GPU or CPU. If it shows CPU even though you have a compatible card, update the driver and restart Ollama.
Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.