How to Set Up Ollama MLX on Mac: Install, Run, and Manage Local AI Models

Has Ollama felt sluggish on your Apple Silicon Mac? There’s a reason. It was running on a general-purpose engine that never fully used your hardware. MLX changed that. It’s Apple’s own machine learning framework, built just for M-series chips. Update Ollama, and if your Mac qualifies, models load faster and use less memory. No extra setup needed.

This guide skips the “what is Ollama” basics; you already know that part. What you need: turning on MLX, picking model sizes that fit your Mac’s memory, and connecting Ollama to your own apps and scripts.

What Changed With Ollama’s MLX Backend

MLX is Apple’s machine learning framework. It’s built around the unified memory architecture in M-series chips. The CPU and GPU share one pool of RAM, instead of copying data back and forth between separate pools. Starting with Ollama v0.19 (March 2026), MLX became an alternative to the older general-purpose engine, which was based on llama.cpp. It runs on Apple Silicon Macs only. Ollama has kept shipping MLX-related updates in the releases since then — as of September 2026 the stable line is up in the 0.34.x series — so don’t treat v0.19 itself as today’s target. What matters for this guide is being on 0.19 or newer, not on 0.19 specifically.

In practice, that means faster responses and lower memory use when MLX is active, at least according to Ollama’s own benchmarks. Nothing about how you use Ollama changes. The ollama run command works exactly like it always has. Ollama just picks a faster engine under the hood when your Mac qualifies.

One caveat worth knowing upfront: MLX launched as a preview in March 2026, and Ollama has continued expanding its engine and model support in the releases since — this isn’t a single fixed feature, so expect the details (which models qualify, how memory is handled) to keep shifting release to release. It still only activates automatically on Macs with 32GB or more of unified memory. Got 16GB or less? Ollama still works fine. It just falls back to the older Metal-based engine, still capable, just not as quick.

Before You Begin

Make sure you have:

• An Apple Silicon Mac (M1, M2, M3, M4, or later) for MLX acceleration
• macOS 13 (Ventura) or later to run Ollama at all — the MLX backend specifically is built and tested against macOS 14 (Sonoma) and macOS 15 (Sequoia), so update to 14 or newer if you’re chasing MLX performance specifically
• At least 16GB of unified memory (32GB+ needed to enable MLX itself)
• 10–40GB of free disk space, depending on which models you plan to download
• Comfort opening and typing commands in Terminal
• Homebrew installed (optional, but makes updates easier); get it at brew.sh

Requirement	Details
Device	Mac with Apple Silicon (M1 or later)
OS version	macOS 13+ to run Ollama; macOS 14 (Sonoma) or 15 (Sequoia) for the MLX backend specifically (check in Apple menu > About This Mac)
Memory for MLX	32GB+ unified memory to activate MLX automatically
Memory minimum	16GB to run smaller models with the standard engine
Disk space	10–40GB free, depending on model sizes

Step-by-Step Guide

Step 1: Check Your Mac’s Chip and Memory

Before installing anything, confirm your Mac qualifies for MLX acceleration. Click the Apple menu > About This Mac.

Look at two fields:

• Chip: Should read M1, M2, M3, M4, or a later Apple Silicon chip, not “Intel”
• Memory: The amount of unified memory installed (8GB, 16GB, 32GB, 64GB, etc.)

Write this number down. It decides whether MLX turns on and which model sizes fit your setup.

Tip: Shopping for a new Mac to run local AI models well? Memory matters more than almost anything else, even more than chip generation. A 32GB M1 will outrun a 16GB M4 for this job because model size is capped by available RAM.

Step 2: Confirm Your macOS Version

While you’re in About This Mac, check the macOS version under your Mac’s name. It looks something like “macOS Sonoma 14.6” or “macOS Sequoia 15.1.” Ollama itself runs on macOS 13 (Ventura) or later, but the MLX backend specifically is built and tested against macOS 14 (Sonoma) and macOS 15 (Sequoia) — that’s what you want if you’re chasing MLX performance.

If you’re behind, go to System Settings > General > Software Update and install any pending updates before continuing.

Step 3: Install Ollama on macOS

You have two options: Homebrew (faster to update later) or the direct installer.

Option A: Homebrew:

brew install ollama

Option B: Direct download:

Go to ollama.com/download and download the macOS installer. Open the downloaded file and drag the Ollama icon into your Applications folder, just like installing any other Mac app.

Step 4: Install Ollama on Windows

Setting up Ollama on a Windows PC too, maybe a work laptop alongside your Mac? The process is simpler, but it skips MLX entirely. MLX is Apple Silicon-only.

• Go to ollama.com/download in your browser and download the Windows installer
• Double-click the downloaded .exe file
• Follow the on-screen prompts to complete installation
• Ollama installs itself as a background service, accessible from Command Prompt, PowerShell, or the Ollama desktop app

On Windows, Ollama uses its standard engine instead of MLX. You still get full local AI features. You just don’t get the Apple Silicon speed boost this guide is built around.

Step 5: Launch Ollama and Verify Your Version

On your Mac, open the Ollama app from Applications. It runs quietly in the background; look for its icon in the menu bar. If you’d rather work entirely from the command line, start the server manually instead:

ollama serve

Leave that Terminal window running, or just let the background app handle it. Either works fine. Next, confirm which version you have installed:

ollama --version

You should see output similar to this (the exact number will be newer for you — Ollama has shipped many releases since 0.19; what matters is that it reads 0.19 or higher, not this specific build):

ollama version is 0.19.4

Tip: MLX support requires version 0.19 or later. If your version number starts with 0.18 or lower, move on to the next step to update.

Step 6: Update Ollama for MLX Support

If you installed Ollama a while back, update it now. This is the step that actually turns MLX on.

On macOS with Homebrew:

brew upgrade ollama

On macOS with the direct installer: Download the latest installer package from ollama.com and reinstall. It overwrites the older version automatically.

On Windows: Re-download and run the installer from ollama.com; it detects and replaces the existing installation.

Run ollama --version again afterward to confirm you’re on 0.19 or later.

Step 7: Pull and Run Your First Model

Now for the fun part. Pick a model size that fits your Mac’s memory. Check the recommendations table below, then run it with one command. For a 16GB Mac, a good starting point is Llama 3.1’s 8B model:

ollama run llama3.1:8b

The first time you run this, Ollama downloads the model automatically. You’ll see a progress bar as layers get fetched and verified. Once it finishes, you land in an interactive chat prompt right in Terminal. Type a question and press Return:

>>> Explain unified memory in one sentence.

To exit the chat session, type /bye or press Ctrl + D.

Tip: The first response after loading a model is usually slower than the ones after it. That’s just the model loading into memory, not a performance problem.

Step 8: Try the Ollama Desktop App

If you’d rather skip Terminal for everyday chatting, open the Ollama app directly. It gives you a graphical chat window, similar to ChatGPT. There’s a model picker at the top, and you can drop in files or images for multimodal models.

Click the model dropdown, select a downloaded model, and start typing in the chat box at the bottom.

Step 9: List Your Downloaded Models

As you experiment, you’ll accumulate several models; it happens faster than you’d think. See what’s installed and how much space each one takes:

ollama list

Expected output looks like this:

NAME ID SIZE MODIFIED
llama3.1:8b a1b2c3d4e5f6 4.9 GB 2 hours ago
qwen2.5:14b f6e5d4c3b2a1 9.0 GB 1 day ago

Step 10: Switch Between Downloaded Models

You don’t need to uninstall anything to switch models; just run a different one by name:

ollama run qwen2.5:14b

Ollama loads the new model into memory and unloads the old one automatically. You land in a fresh chat session. This is handy for comparing how different models handle the same prompt, say, testing a coding model against a general-purpose one.

Step 11: Remove Models You Don’t Need

Larger models eat up tens of gigabytes fast. Clean up ones you’re not using anymore:

ollama rm qwen2.5:14b

Run ollama list again afterward to confirm it’s gone. Freed disk space comes back immediately.

Step 12: Confirm MLX Is Actually Active

While a model is running, open Activity Monitor (find it via Spotlight with Cmd + Space, then type “Activity Monitor”). Click the Memory tab and look for the ollama process.

With MLX active, memory usage should track close to the model’s stated size. Responses should also feel noticeably snappier than older Ollama versions on the same hardware. If memory usage looks about the same as before your update, jump to the troubleshooting section.

Configuration: Choosing the Right Model Size for Your Mac

Model size and your Mac’s memory need to match. Get it wrong, and you’ll see sluggish performance, or the model won’t load at all. Here’s a realistic breakdown:

Mac Memory	Recommended Model Size	MLX Status	Example Models
16GB	7B–8B parameters, quantized	Standard engine (MLX not active in current preview)	llama3.1:8b, mistral:7b, gemma2:9b
32GB	12B–14B parameters	MLX active	qwen2.5:14b, gemma2:27b (quantized)
64GB	30B+ parameters	MLX active, with headroom	llama3.1:70b (quantized), qwen2.5:32b

Heads up: Treat these tiers as a starting point, not a hard rule. Quantization level, context length, model family, and whatever else is competing for memory in the background can all shift what actually fits — the same “32GB tier” model can run fine on one setup and choke on another depending on those factors.

Note: “B” refers to billions of parameters. Roughly, that’s the model’s complexity and knowledge capacity. Bigger doesn’t automatically mean better. A well-tuned 8B model can beat a poorly-suited 30B model at something simple, like summarization. It’ll also use far less memory.

Browse available models and their size tags at ollama.com/library before downloading. Each listing shows quantization options; compressed versions that trade a little accuracy for a smaller footprint.

Performance Expectations by Chip

Real-world speed depends on your chip generation, memory bandwidth, and parameter count:

Chip	Typical Experience with MLX-Eligible Setup
M1 / M1 Pro / M1 Max	Solid for 8B–13B models with 32GB+ RAM; noticeably faster with MLX vs. the older engine
M2 / M2 Pro / M2 Max	Comfortable with mid-size models; faster token generation than M1 at the same model size
M3 / M3 Pro / M3 Max	Strong across 13B–30B models; early user reports and Ollama’s own benchmarks suggest MLX gains are often most noticeable here, though no rigorous independent per-chip benchmarks are published
M4 / M4 Pro / M4 Max	Among the fastest in the current Apple Silicon lineup for this workload; handles 30B+ models fluidly on 64GB configurations

Response speed is usually measured in tokens per second, roughly, how fast words appear on screen. With MLX active, chat responses feel conversational. Older local AI setups tend to feel slower and choppier by comparison.

Setting Up Ollama’s Local API

Ollama runs a local web server on port 11434 that speaks an OpenAI-compatible format. That means you can connect it to your own scripts, note-taking apps, or coding tools instead of chatting through Terminal.

With Ollama running in the background, test the API with a simple request:

curl http://localhost:11434/api/generate -d '{
  "model": "llama3.1:8b",
  "prompt": "Write a one-sentence product description for a <a href="https://www.amazon.com/dp/B0DGJ4QQ5W?tag=switchingtomac-20" target="_blank" rel="noopener noreferrer">MagSafe charger</a>.",
  "stream": false
}'

You’ll get back a JSON response with the generated text. For a quick, readable view, pipe it through Python’s built-in JSON module as a separate step instead of chaining commands. Or just open the response in a text editor if you’d rather not parse JSON by hand.

For app integrations that expect an OpenAI-style API, Ollama also exposes a compatible endpoint:

curl http://localhost:11434/v1/chat/completions -d '{
  "model": "llama3.1:8b",
  "messages": [{"role": "user", "content": "Summarize this in two sentences."}]
}'

That means plenty of tools built for OpenAI’s API can point at http://localhost:11434 instead of a cloud endpoint. Think code editors with AI chat panels, browser extensions, or note apps. Often, that’s just changing a base URL setting and leaving the API key field blank or filled with a placeholder.

If you’re building something in Python, install Ollama’s official client library instead of hand-rolling requests:

pip install ollama

import ollama

response = ollama.chat(model='llama3.1:8b', messages=[
    {'role': 'user', 'content': 'What is unified memory?'}
])
print(response['message']['content'])

This is a solid starting point for a personal writing assistant or a local research tool. You can bolt AI features onto an existing script without paying per-token cloud costs.

Tips and Troubleshooting

Why isn’t MLX activating on my Mac?

Why it happens: MLX is still in preview, and it only activates automatically on Macs with 32GB or more of unified memory. Macs with 16GB or 8GB keep using the older engine. That’s expected, nothing to worry about.

Fix: Check your memory via Apple menu > About This Mac. Below 32GB, MLX won’t turn on yet for your hardware. If you do have 32GB or more, confirm you’re running Ollama 0.19 or later with ollama --version. Then restart the Ollama app, or run ollama serve fresh.

Model loading is slow or times out

Why it happens: Loading large models from a slow external drive can cause GPU timeouts. Ollama’s recent updates specifically fixed this.

Fix: Update to the latest Ollama version. Store your models on your Mac’s internal SSD rather than an external SSD or USB-C hub-connected drive when you can.

My Mac feels sluggish while a model is running

Why it happens: Larger models (30B parameters and up) can eat most of your available unified memory. That leaves little for anything else you’re running.

Fix: Drop to a smaller model that fits your RAM tier (see the table above). Close unused apps, and check Activity Monitor > Memory for memory pressure. Red means you’re overcommitted.

The desktop app doesn’t match system dark/light mode

Why it happens: Some users have reported this on older Ollama app builds; Ollama doesn’t publish a per-version changelog entry confirming exactly which releases were affected.

Fix: Update Ollama to the latest version — most reports of this resolve after updating. Quit and reopen the app.

Multiple Ollama windows open unexpectedly

Why it happens: Older versions have sometimes launched a second instance instead of reusing the one already running.

Fix: Update to the latest version — this has been reported fixed in current releases, and only one instance should launch.

Tips and Hidden Features

• Run models fully offline: Once downloaded, models work with zero internet connection, handy on flights or anywhere Wi-Fi is unreliable.
• Check model info before downloading: Run ollama show llama3.1:8b to see a model’s parameter count, context length, and license before committing disk space.
• Free up space fast: Run ollama list, then ollama rm on anything you haven’t used in a month. Models add up quickly on a laptop’s internal SSD.
• Use an external SSD for extra headroom: Experimenting with many large models on a 512GB Mac? A Thunderbolt dock with an external SSD adds storage without touching your boot drive’s free space. Just avoid running the largest models directly off it; see the troubleshooting note above.
• Multitask smart: Running a 30B+ model alongside video editing or heavy browser tabs competes for the same unified memory pool. Close what you don’t need first.

Wrapping Up

Update to Ollama 0.19 or later, and confirm you’re on a 32GB+ Apple Silicon Mac. MLX just works in the background, no toggle to flip, no setting to remember. The bigger decision is picking model sizes that respect your Mac’s memory ceiling. That should feel like second nature by now.

This update makes local AI on a Mac feel genuinely competitive with cloud chatbots for everyday tasks. The credit goes to software finally using Apple’s hardware well, not to any sudden leap in the models themselves.

Step	Action	Applies To
1–2	Check chip, memory, and macOS version	macOS
3–4	Install Ollama	macOS, Windows
5–6	Verify and update to 0.19+	macOS, Windows
7–8	Pull and run your first model	macOS, Windows
9–11	List, switch, and remove models	macOS, Windows
12	Confirm MLX is active via Activity Monitor	macOS

Resources

• Ollama MLX blog announcement
• Ollama MLX performance update
• Ollama new desktop app announcement
• Ollama GitHub releases (version history)
• Ollama model library