Ollama tutorial: install and run a local model

To use Ollama, install the application for your operating system, then run a local model with the command provided in this tutorial. The first run downloads the model; subsequent conversations run on your computer, provided you keep using a local model and do not enable a remote service.

1. Prepare your computer and choose a starting model

Before installing, make sure you have an internet connection to download the software and models, available disk space, and a terminal. On Windows, use PowerShell; on macOS or Linux, open your Terminal application.

Memory requirements depend on the model, its compression, and the length of your conversations. The download size does not account for all the memory a model uses during a chat. Avoid choosing your first model based solely on how much free disk space you have.

This tutorial uses Gemma 3, with the specific identifier gemma3:1b. This text variant appears in the official Ollama library, consulted for this October 2026 update. Its purpose here is to check that the installation works, rather than to suggest it is the best choice for every task.

The suffix after the colon identifies a model variant. Keep the same identifier throughout the commands: changing the suffix can trigger another download and alter the hardware requirements.

If you are planning a more demanding setup, consult the guide to running LLMs locally. It explains how hardware, model size, and intended use fit together. Quantization, which reduces the numerical precision of model weights, also helps explain why variants of the same model can take up different amounts of space.

2. Install Ollama on your operating system

Follow only the instructions for your computer. The requirements below come from the official documentation consulted in October 2026; check them again before installing on an older machine.

On Windows

The Windows documentation specifies Windows 10 22H2 or later. Download the installer from the official website linked in that documentation, then run it.

The standard installation does not require administrator privileges. Once installed, Ollama runs in the background, and its command becomes available in PowerShell or another terminal.

Open a new PowerShell window after installation. This lets the terminal pick up the configuration changes. You can then move on to the version check below.

On macOS

The macOS documentation requires macOS Sonoma 14 or later. It lists CPU and GPU support on Apple silicon Macs, and CPU-only support on x86 machines.

Download the official disk image, open it, and drag the application into the Applications folder. Then launch Ollama. If the application offers to install the link that makes the command available in your terminal, complete that step.

On Linux

The Linux documentation provides this installation command:

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

It downloads a script and then executes it. On a work computer, have this method approved according to your organization’s rules.

Next, check that the command is installed:

ollama -v

This check also applies to Windows and macOS. If the terminal displays a version, the command is available. Make a note of that version in case you need to troubleshoot later; you do not need a particular version number to follow this tutorial.

3. Download the model and start a conversation

Run this command in your terminal:

ollama run gemma3:1b

On the first run, wait for the download and loading process to finish. If the model is already on your computer, Ollama may be able to use it immediately. Do not confuse this initial wait with the time needed to generate each response.

When the chat prompt appears, enter the following request, which asks for a professional rewrite in French:

Réponds en français. Reformule ce message dans un ton professionnel,
sans ajouter d'information :
La réunion est reportée à jeudi. Le lieu reste à confirmer.

Check that the response preserves both pieces of information: the meeting has been postponed until Thursday, and the location is still to be confirmed. It should not invent a time or an address. This test checks both that the model runs and that it can perform a useful task: staying faithful to supplied text.

Continue in the same conversation with this request to shorten the wording while keeping the uncertainty about the location explicit:

Raccourcis ta proposition et conserve explicitement
l'incertitude sur le lieu.

The model can use the earlier exchanges in this session. That does not mean it knows about your other conversations or has read your documents.

To end the conversation, enter:

/bye

The command-line reference also covers downloading and managing models. To download a model without immediately starting a conversation, use:

ollama pull gemma3:1b

Keep a small set of representative requests: rewriting, summarizing supplied text, and extracting information. You can reuse these same requests to evaluate another model, rather than judging it by a single successful response.

4. Check local execution and manage resources

Ollama also offers cloud features. Having the application installed on your computer is therefore not enough to establish where every request is processed.

To disable Ollama’s cloud features, create or edit the server.json file in the .ollama folder in your home directory. The official FAQ documents this setting:

{
  "disable_ollama_cloud": true
}

If the file already contains settings, add this property while keeping the JSON valid. Restart Ollama after making the change.

While a model is generating a response, open another terminal and run:

ollama ps

The PROCESSOR column shows whether the loaded model uses the CPU, GPU, or both. This can help explain slow performance; it does not measure response quality.

It also helps to distinguish these operations:

Operation	Command	Purpose
List downloaded models	`ollama ls`	Find the model identifiers available locally
List loaded models	`ollama ps`	See which resources are in use
Unload the model	`ollama stop gemma3:1b`	Free the memory occupied by this model
Delete the model	`ollama rm gemma3:1b`	Remove its local copy

Unloading and deleting serve different purposes. After deleting a model, you may need to download it again before you can use it.

For a practical check, download the model, temporarily disconnect from the network, and try another text request. This confirms that the workflow works offline. The check does not cover extensions or applications you might add later.

5. Call the model from a program

This step is optional: the terminal is enough for chatting. The local API becomes useful when you want to add text generation to a script or application.

The following example adapts the endpoint described in the chat API documentation. Run it in Bash on macOS or Linux while Ollama is running:

curl http://localhost:11434/api/chat \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma3:1b",
    "messages": [
      {
        "role": "user",
        "content": "Reformule en français professionnel : le lieu reste à confirmer."
      }
    ],
    "stream": false
  }'

The address localhost refers to your own computer. The model field contains the identifier of the downloaded model; messages contains the text you send. Disabling stream returns a complete response instead of a sequence of fragments.

The result is a JSON document. The generated text is in message.content. In an application, handle the model’s response separately from connection errors or failures to load the model.

An independent API call does not automatically include previous exchanges: your program must send whatever conversation history is needed. Include only relevant material, because longer conversations use more resources.

For this first setup, keep API access confined to your computer. Making an API available over the network requires additional configuration and access controls. Local execution does not replace those protections.

6. Common problems

The command cannot be found

Close the terminal, open a new one, and try the version check again. On macOS, check that the application has finished setting up the terminal command. If the problem persists, return to the official installation instructions for your operating system.

The terminal cannot connect to Ollama

Check that the application is running. On Linux, if no service is already active, start the server in a terminal:

ollama serve

Keep that window open and launch the model from another terminal. If a message says the address is already in use, look for the existing process before trying to start another instance.

The download fails

Check your connection and available disk space, then retry the download. On a corporate network, a proxy or filtering policy may block access. Have that issue investigated without disabling network protections.

Generation is slow or runs out of memory

Close applications that use a lot of resources and check which models are loaded. Try a fresh conversation with a short instruction. If you selected a larger variant, return to the model used in this tutorial.

The guide to choosing a model can then help you weigh response quality, speed, and hardware constraints.

The answer is wrong even though installation succeeded

A working installation only establishes that the model runs. It does not prevent a hallucination.

Check responses against the original documents. To compare models, consult the language model profiles and the benchmark lab, then check how they perform on your own requests.

Checklist before regular use

The first goal is a repeatable workflow: launch the same model, give it a familiar request, and check its response. Expand your use gradually, keeping examples of observed errors in your evaluation set.

The version command works in a new terminal.
Your chosen model is downloaded, and you have kept its identifier.
A simple request produces a response you have checked.
Local execution is confirmed, and cloud features are disabled if needed.
You understand the difference between downloading, loading into memory, and deleting.
Responses involving sensitive matters receive human review.

For more detail on hardware constraints, the article Running an LLM locally builds on this tutorial. If you are still deciding which software to use, see the LM Studio and Ollama comparison.

Frequently asked questions

Do I need an account to use Ollama locally?

The local setup covered here requires neither an account login nor an API key. If an interface asks you to sign in, check whether you have selected a cloud feature instead of your downloaded model.

Can Ollama run without a graphics card?

It can run on a CPU, but speed depends on the model and your hardware. Starting with a small model lets you check whether response times are acceptable for your needs.

Can I use Ollama without an internet connection?

Yes, provided the local model is already downloaded and the task does not depend on an external service. Install any other software you need before disconnecting, too: installation and new downloads require a connection.

Does installing Ollama let me analyze all my documents?

No. A model does not automatically know about the files on your computer. You need to provide the relevant content or configure an application that can prepare and search your documents.

Ollama step by step: install and run a model locally