Setting up a language model to run locally on Windows or Mac once meant navigating Python environments, CUDA driver versions and a tangle of dependencies. Ollama collapses all of that into a single installer on both platforms, and a few minutes after the download finishes you have a local model server answering requests on port 11434 with nothing leaving your PC.

Quick Answer

Download the native installer from ollama.com, run it like any other app, and Ollama starts a background service listening on http://localhost:11434. On Windows it runs natively (no WSL2 needed) and adds a system tray icon; on Mac it installs a menu bar app and desktop client. Then run one command, ollama run, to pull and chat with your first model.

Installing Ollama on Windows

  1. Download the installer. Go to ollama.com/download and grab the Windows .exe. Ollama has shipped native Windows support since version 0.3, so you no longer need WSL2 or Linux underneath it.
  2. Run the installer. Double-click the .exe and follow the prompts. It installs like a normal Windows application, sets itself to start automatically, and adds an icon to the system tray.
  3. Confirm the service is up. Open Command Prompt or PowerShell and run ollama --version. The background service is already listening on localhost:11434 at this point.
  4. Prefer the command line? If you use a package manager, winget install Ollama.Ollama does the same job from a terminal.

Installing Ollama on Mac

  1. Download the macOS installer from the same ollama.com download page, or use Homebrew with brew install ollama if you already run it.
  2. Open the app. The installer adds a menu bar icon and the native desktop application that Ollama shipped in 2025. Launching it starts the local server in the background.
  3. Check it from Terminal with ollama --version, exactly as on Windows. The API endpoint is identical: http://localhost:11434.

Pulling and running your first model

With Ollama installed, the workflow is the same on both platforms. In a terminal, run something like ollama run llama3.2 and Ollama downloads the model weights once, then drops you into an interactive chat. Compact models in the 3B to 8B class run adequately on everyday hardware; larger models benefit from more RAM and a capable GPU. If you need a rig for this kind of sustained inference work, the AI-ready PCs at Evetech are specced around the GPU memory and system RAM that local model serving actually depends on.

Understanding port 11434

The number worth remembering is 11434. That is the default port for Ollama's built-in REST API, and any compatible client, from a VS Code extension to a custom script, reaches the model by talking to http://localhost:11434. If you want other devices on your home network to use the same model server, set the environment variable OLLAMA_HOST=0.0.0.0:11434 so it listens on all interfaces rather than localhost only. Leave it on localhost if you only ever call it from the same machine.

GPU acceleration across platforms

Ollama handles GPU acceleration differently depending on your hardware.

NVIDIA (Windows and Mac). Any NVIDIA GPU with CUDA compute capability 5.0 or higher -- the GTX 900 series and newer -- works out of the box. Ollama detects it automatically and routes inference through CUDA for fast token generation.

AMD (Windows). As of 2026, AMD GPU acceleration via ROCm is supported on Linux only. On Windows, an AMD card falls back to CPU inference, which is noticeably slower. If you are on Windows and want GPU-accelerated local models, an NVIDIA card is the current practical choice.

Apple Silicon. M1, M2, M3 and M4 Macs are fully supported and require no configuration. Starting with Ollama 0.19, released in early 2026, inference on Apple Silicon moved from llama.cpp to Apple's MLX framework, which Ollama's own benchmarks show is roughly 1.6 to 2 times faster on decoding. Install Ollama on an M-series Mac and Metal acceleration is active immediately.

What hardware actually matters

Model size is governed mostly by RAM and GPU VRAM, not raw clock speed. An 8B model needs roughly 8GB or more of available memory to run smoothly, and an NVIDIA GPU dramatically speeds up generation compared to running on the CPU alone. Practical guidance: 8GB of system RAM handles a 7B model, 16GB handles a 14B, and 24GB or more gets you into the 32B range. Popular configurations that balance these specs are listed in the best-selling desktop PCs at Evetech, which give a sense of what people pair with this kind of workload.

Frequently Asked Questions

Do I still need WSL2 to run Ollama on Windows?

No. Ollama has had native Windows support since version 0.3, so the .exe installer runs it directly on Windows without WSL2 or a Linux layer.

What port does Ollama use?

Port 11434 by default. The local REST API lives at http://localhost:11434, and that is the address every compatible client connects to.

How do I let other devices on my network reach Ollama?

Set the environment variable OLLAMA_HOST to 0.0.0.0:11434, which tells the server to listen on all network interfaces instead of localhost only. Then other machines can reach it at your PC's local IP on that port.

How much RAM do I need to run a model?

It scales with model size. A small 3B model runs on very modest hardware, while an 8B model wants roughly 8GB or more of available memory. Larger models benefit from a GPU with plenty of VRAM.

Does Ollama start automatically?

Yes. After installation the service runs in the background on startup, listening on localhost:11434, so the model server is ready whenever you open a terminal or client.

Want a machine that handles local models without breaking a sweat? Browse the AI-ready PC range at Evetech and pick a build with the GPU memory and RAM that local inference really depends on.