Setup guide

Ollama on Mac in 2026: setup, and why it now runs on MLX

By Claudio A. G. M. Filho · Updated October 21, 2026

Ollama wraps local-model execution in an app, CLI, and HTTP API. For direct MLX use without this wrapper, see our mlx-lm guide. The big 2026 change: since v0.40.0 (released September 25, 2026), model architectures supported by Ollama's MLX runtime run on MLX automatically on Apple Silicon; do not assume that every model or tag selects the same backend.

Install

Download the macOS app from ollama.com/download, drag it to Applications, and launch it once. Then in Terminal:

ollama --version

Choose an exact model identifier from the official model library, check its supported runtime and license, and estimate its memory needs using our calculator. Downloading model weights may use many gigabytes of storage and network transfer. The commands on this page are instructions for readers; MacModelFit has not executed or benchmarked this setup.

Run a model

This exact Llama 3.1 8B tag exists in Ollama’s library. Its listed download is about 4.9 GB; download size is not total runtime memory. Confirm the license and available memory first.

ollama run llama3.1:8b

The first run downloads the model, then opens a chat. Use /bye to leave it. This tag alone does not prove which runtime backend is selected.

Check which engine you got

ollama ps

The official FAQ documents a PROCESSOR column showing where the model is loaded. That is not a speed benchmark or, by itself, proof of the backend. Consult the installed version's release notes and logs to confirm backend selection; output columns can change.

Why MLX matters

Ollama introduced its MLX engine in preview on March 30, 2026, saying it takes advantage of unified memory and, on M5-family chips, the GPU Neural Accelerators to speed up time-to-first-token and generation. Its preview announcement also describes NVFP4 support. Quantization quality depends on the model and evaluation; this guide does not establish that one format is universally better.

Use it from other apps

Ollama serves a local HTTP API on localhost:11434. Coding agents and chat apps can point at it; Ollama also offers ollama launch to start supported agents with a chosen model.

Tips

• Models unload after a few minutes idle by default; set OLLAMA_KEEP_ALIVE to change that.
• Check the actual model metadata and installed runtime; a name suffix is not a universal backend contract.
• Confirm the runtime's current macOS and hardware requirements before downloading a model.