How to Install and Run Ollama on Windows for Local AI

Ollama makes it possible to run large language models directly on a Windows PC without sending every prompt to a cloud AI service.

You can use it to experiment with models, build local AI applications, connect AI to development tools, or simply run a private chatbot on your own computer.

And on current versions of Ollama, Windows users do not need to install WSL2 just to get started. Ollama provides a native Windows installation path.

This guide explains how to install and run Ollama locally on Windows, choose an appropriate model, verify whether your GPU is being used, connect to the local API, and troubleshoot common problems.

What Is Ollama?

Ollama is a tool for downloading, running, and managing supported AI models locally.

Instead of sending each prompt to a remote API, the model runs on your computer using available CPU, memory, and supported GPU resources.

Ollama also exposes a local API, which makes it possible to connect local models to:

• Python applications;
• development tools;
• automation workflows;
• local chat interfaces;
• RAG applications;
• other software that can send HTTP requests.

Why Run AI Locally?

Local AI has several practical advantages.

More Control Over Data Flow

When inference runs locally, prompts can remain on the machine rather than being sent to a remote inference API.

However, local execution does not automatically make an entire workflow private.

If your application also connects to cloud storage, web search, remote APIs, telemetry services, or other external systems, those parts of the workflow may still transmit information.

No Per-Request API Charge for Local Inference

Once a model has been downloaded, local inference does not require paying a hosted model provider for each prompt.

You still pay indirectly through hardware, electricity, storage, and your own maintenance time.

Works Without Continuous Internet Access

After Ollama and the required model files are available locally, many basic inference tasks can run without an active internet connection.

Useful for Development

A local model is convenient when testing:

• AI application prototypes;
• prompt workflows;
• local APIs;
• RAG systems;
• automation experiments.

What Does Your Windows PC Need?

There is no single hardware specification that guarantees a good Ollama experience.

Requirements depend heavily on:

• the model;
• model size;
• quantization;
• context length;
• available RAM;
• available GPU memory;
• how fast you expect inference to run.

RAM Matters

Smaller models are generally easier to run on ordinary laptops.

Larger models require substantially more memory.

If you are new to local AI, start with a small model rather than downloading the largest option immediately.

Disk Space Matters Too

Model files can consume several gigabytes each.

For example, the current Ollama library lists the default Llama 3.2 model at about 2 GB, while the current Phi-4 package is about 9.1 GB. Exact sizes depend on the specific model variant.

GPU Acceleration Is Optional

Ollama can run models using CPU resources, but supported GPUs can make inference significantly more practical.

Whether a particular GPU is supported and how Ollama uses it can change as Ollama evolves, so check current official documentation rather than relying on an old CUDA version requirement.

Step 1: Install Ollama on Windows

The simplest current option is the official Windows installer.

Visit the official Ollama Windows download page:

Download Ollama for Windows

Ollama currently supports Windows 10 or later.

The official page also provides a PowerShell installation command:

irm https://ollama.com/install.ps1 | iex

If you prefer not to execute an installation script directly from PowerShell, use the downloadable Windows installer from the official site instead.

Verify the Installation

Open PowerShell or Windows Terminal and run:

ollama --version

If Ollama is installed correctly, the command should return the installed version.

Step 2: Run Your First Model

The basic syntax is:

ollama run MODEL_NAME

Ollama downloads the model automatically if it is not already available locally.

Beginner-Friendly Example: Llama 3.2

Ollama’s current library provides Llama 3.2 in relatively small 1B and 3B variants, making it a practical starting point for users who do not have a high-end workstation.

Run:

ollama run llama3.2

The current default package is approximately 2 GB and uses the 3B model variant.

For an even smaller version:

ollama run llama3.2:1b

Another Option: Phi-4

Phi-4 is a substantially larger model.

Ollama currently lists Phi-4 as a 14B-class model with a package size of approximately 9.1 GB and a 16K context window.

Run:

ollama run phi4

Because Phi-4 is much larger than Llama 3.2 1B or 3B, expect it to require more memory and generally more capable hardware.

Mistral

Mistral models are also available in Ollama’s model library.

For example:

ollama run mistral

Before choosing a model, check its current Ollama library page for size and available variants.

Step 3: Chat With the Model

After running:

ollama run llama3.2

you will enter an interactive prompt.

You can type something like:

Explain the difference between RAM and storage in simple English.

The response is generated by the locally running model.

Exit the interactive session using the appropriate Ollama command or terminal shortcut shown by the current CLI.

Step 4: See Which Models Are Installed

Run:

ollama list

This displays models currently stored locally.

It is useful when your model library begins consuming significant disk space.

Step 5: Download a Model Without Starting a Chat

You can download a model separately using:

ollama pull llama3.2

This is useful when you want to prepare a machine before working offline.

Step 6: Remove a Model You No Longer Need

Local models can occupy a lot of storage.

Remove an unused model with:

ollama rm MODEL_NAME

For example:

ollama rm llama3.2

Run ollama list afterward to confirm which models remain installed.

Step 7: Check Whether Your GPU Is Being Used

GPU detection depends on your hardware and current Ollama support.

For an NVIDIA system, one useful Windows diagnostic tool is:

nvidia-smi

Run it while a model is generating output and observe whether Ollama-related processing appears and GPU memory usage changes.

If GPU acceleration does not appear to work:

• update the GPU driver;
• restart Ollama;
• restart Windows if needed;
• verify that the GPU is supported by the current Ollama release;
• check current Ollama troubleshooting documentation.

Avoid installing random CUDA versions simply because an older tutorial recommends one. Current native Ollama releases may manage GPU support differently from older WSL-focused setups.

Step 8: Use Ollama’s Local API

One of Ollama’s most useful features is its local HTTP API.

The default local endpoint commonly uses:

http://localhost:11434

For example, the current Ollama model library demonstrates chat requests using the local API.

Example With curl

curl http://localhost:11434/api/chat ^
  -d "{\"model\":\"llama3.2\",\"messages\":[{\"role\":\"user\",\"content\":\"Hello!\"}]}"

Command quoting differs between Windows shells, so adjust the syntax if you use PowerShell instead of Command Prompt.

Step 9: Use Ollama With Python

Ollama provides a Python package that can communicate with the local service.

Install it:

pip install ollama

Then create a simple example:

from ollama import chat

response = chat(
    model='llama3.2',
    messages=[
        {
            'role': 'user',
            'content': 'Explain local AI in three sentences.'
        }
    ]
)

print(response.message.content)

This gives you a simple foundation for building local AI applications without manually constructing HTTP requests.

Step 10: Use Ollama With Other Frameworks

Ollama can also be used as a local model backend for software that supports it.

Depending on the project, that may include:

• LangChain;
• LlamaIndex;
• local chat interfaces;
• development assistants;
• automation platforms.

Integration packages and APIs change quickly, so use the current documentation for the framework you choose rather than copying an older package name.

Step 11: Create a Custom Model Configuration

Ollama supports a Modelfile for defining model behavior and parameters.

A simple example might look like:

FROM llama3.2

SYSTEM """
You are a concise technical assistant.
Explain concepts using simple examples.
"""

PARAMETER temperature 0.3

Then create the model:

ollama create my-assistant -f Modelfile

Run it:

ollama run my-assistant

This is useful when you want reusable instructions rather than repeatedly entering the same system prompt.

Step 12: Be Careful With Context Size

Long context windows allow models to work with more text at once, but larger contexts consume additional memory and can reduce performance.

Do not automatically set the context to the largest number supported by a model.

Increase it only when your task actually requires more context.

The maximum supported context also depends on the specific model, not just Ollama itself.

Do You Still Need WSL2?

Not for a normal Ollama installation on current Windows systems.

The native Windows installer is usually the simplest starting point.

WSL2 may still be useful if:

• your development environment already runs primarily in Linux;
• you need Linux-specific tooling around Ollama;
• your application stack is deployed or tested inside WSL;
• you specifically prefer Linux package and shell workflows.

But WSL2 should be treated as an optional development environment, not a mandatory Ollama prerequisite.

Choosing a Model for Your Hardware

The easiest approach is to start smaller and move upward.

Use Case	Suggested Starting Point	Why
Basic experimentation	Llama 3.2 1B	Small download and relatively lightweight
General local chat	Llama 3.2	Practical balance for many PCs
More demanding reasoning	Phi-4	Larger model, higher hardware requirements
Experimenting with alternatives	Mistral or another library model	Useful for comparing behavior

This is not a universal performance ranking.

Model quality and speed depend on the task, hardware, model version, quantization, and context settings.

Troubleshooting: “ollama” Is Not Recognized

If PowerShell returns an error similar to:

ollama : The term 'ollama' is not recognized...

try:

1. close the terminal;
2. open a new PowerShell or Windows Terminal session;
3. run ollama --version again;
4. verify that Ollama was installed successfully;
5. reinstall from the official Windows installer if necessary.

A newly installed application may not be visible to a terminal session that was already open before installation.

Troubleshooting: Model Is Very Slow

Slow inference can have several causes.

The Model May Be Too Large

Try a smaller model:

ollama run llama3.2:1b

The GPU May Not Be Used

Check GPU activity and update the relevant driver.

The Context May Be Too Large

Reduce the amount of text sent in each request or use a smaller context configuration.

Other Applications May Be Consuming Memory

Close memory-intensive software and try again.

Troubleshooting: Ollama Cannot Download a Model

Check:

• internet connectivity;
• available disk space;
• firewall or proxy restrictions;
• whether the model name is correct;
• whether Ollama services are running normally.

Then retry:

ollama pull llama3.2

Privacy: Is Ollama Completely Offline?

The local model inference itself can run on your machine.

But saying an entire Ollama-based application is automatically “completely offline” would be too broad.

For example, your application may still:

• download models;
• connect to online search;
• call other APIs;
• send telemetry through another application;
• retrieve cloud-hosted documents.

If privacy is important, audit the complete application and network flow—not only the language model runtime.

Security: Be Careful When Exposing the Ollama API

A local API is convenient, but do not expose it broadly to a network without understanding the security consequences.

If another device can reach your Ollama service, that device may potentially interact with the models or applications connected to it.

Before changing host bindings or exposing ports:

• understand which interface Ollama is listening on;
• use firewall controls;
• avoid unnecessary public exposure;
• place authentication or a controlled application layer in front of it where appropriate.

A Simple Beginner Setup

If you just want to experiment with local AI, you do not need a complex stack.

1. Install Ollama for Windows.
2. Open PowerShell.
3. Run ollama run llama3.2.
4. Ask a few questions.
5. Check ollama list.
6. Try the local API when you are ready to build an application.

Only introduce WSL, Docker, LangChain, RAG, or other infrastructure when your project actually requires them.

How This Guide Was Prepared

This guide focuses on the current native Windows installation path and Ollama commands that are documented through Ollama’s official website and model library.

Model availability, package sizes, GPU support, installation methods, context limits, CLI options, and integrations can change as Ollama and individual models are updated.

For that reason, hardware-specific claims and performance numbers in this guide are intentionally presented conservatively rather than as universal benchmarks.

Before troubleshooting a version-specific problem, check the current Ollama documentation, model library, and release information.

Frequently Asked Questions

Can Ollama run directly on Windows?

Yes.

Ollama currently provides a native Windows installer, so WSL2 is not required for a normal installation.

Does Ollama require a GPU?

No.

Models can run using CPU resources, although supported GPU acceleration can substantially improve the experience for many models.

Which model should a beginner try first?

A smaller model is usually the easiest starting point.

For example, Llama 3.2 is currently available in 1B and 3B variants through Ollama.

How much RAM does Ollama need?

There is no universal answer.

Memory requirements depend on the model, quantization, context length, and whether parts of the workload use GPU memory.

If a model is too large for your system, try a smaller variant.

Can Ollama work without internet?

Yes, for local inference after the required software and model files have already been downloaded.

Features that depend on external APIs or online resources still require connectivity.

Can I use Ollama for private company documents?

Running inference locally can reduce exposure to remote inference providers, but privacy depends on the entire system.

Before processing confidential information, check where documents are stored, what integrations are active, what logs are retained, and whether any other component sends information externally.

Where can I find available models?

Use Ollama’s official model library:

Ollama Model Library

Final Takeaway

Running local AI on Windows with Ollama is now much simpler than many older tutorials suggest.

For most users, the process is:

1. install the native Windows version;
2. choose a model appropriate for your hardware;
3. run it with ollama run;
4. verify performance;
5. use the local API when you want to build applications.

You do not need to begin with WSL2, manual CUDA installation, custom quantization, or a complicated Python stack.

Start simple. Confirm that Ollama works. Learn how model size affects your computer. Then add more advanced integrations only when they solve a real problem.