Discover
The right model, in seconds.
Browse thousands of open models from Hugging Face. Each one tells you whether it will run well on your Mac, before you download a thing.
For Mac with Apple silicon
Your own AI.
Right on your Mac.
Download powerful open AI models, make them smaller and faster, shape how they respond, and chat with them in private. Everything happens on your Mac.
Private chatWorks offlineHugging FaceSmaller modelsFaster repliesCustom behaviourModel blendingFine-tuningVisual workflowsSide-by-side testsVoiceImage understandingLocal AI serverShortcuts11 languagesGGUFMLXCore ML
Everything in one app
Find it. Shape it.
Talk to it.
Shrink
Smaller. Faster. Still smart.
Compress a model to a fraction of its size, with a clear report on how much quality it kept.
Customise
Change how it behaves.
Make a model more direct, less preachy, or more like you, without the time and cost of retraining it.
Understand
See how it thinks.
Follow a question as it travels through the model, step by step, in clear and colourful views.
Automate
Build AI workflows visually.
Connect steps on a canvas: sort an incoming message, draft the right reply, read it aloud. No code required.
Chat
Talk privately. Share with your apps.
A fast, private chat for every model you have, plus a built-in server so your other apps can use them too.
0
Conversations sent to the cloud. Your chats and models stay with you.
16GB
Is all it takes. Built to run beautifully on an everyday Mac.
1click
To make a model smaller, faster and lighter on memory.
11
Languages, from English and Japanese to Hindi, Tamil and Telugu.
How it works
From first download
to your own AI.
01 — Discover
Start with a model that fits.
ZMatrix recommends models made for your Mac and tells you up front how each one will run: smoothly, slowly, or not at all.
- Downloads pick up where they left off
- Your Hugging Face account stays safe in your Keychain
02 — Shrink
Make it smaller, without the guesswork.
Create several compact versions at once. ZMatrix compares them for you and recommends the smallest one that still answers well.
- A clear quality report for every version
- Your original model is never touched
03 — Customise
Shape how it answers.
See where a habit like refusing to help lives inside the model, then decide what to change. You stay in control of every edit.
- Cut needless refusals, tone down lecturing, or steer its style
- Blend two models into one
04 — Understand
Look inside, layer by layer.
Watch what the model is thinking at every step on its way to an answer, in views that explain themselves as you go.
- Interactive maps of how the model is built
- Built-in guidance for reading every view
05 — Compare
Know which version is best.
Test how fast and how good each version is, compare answers side by side without knowing which is which, and keep every result.
- Blind side-by-side matchups
- Results you can export and share
06 — Chat
Then simply talk to it.
Chat with any version you've made, with pictures, voice and a live view of its speed and memory.
- Ask two models the same question side by side
- Always know how much memory you're using
See the difference
Scroll to shrink a large AI model to less than a third of its size.
Original
16.4 GB100% of the original · BF16
Approximate sizes for a model with 8 billion parameters. Each square stands for one of the model's numbers. Smaller files use fewer shades, and the red square marks the biggest change.
Smart about memory
It knows what fits.
Before you wait.
ZMatrix checks your Mac's memory before it loads anything. If a model is too big, it tells you why and finds a way that works, instead of slowing your Mac to a crawl.
- Load it as it isWhen there's room, nothing changesfirst
- Use a lighter memory formatA small trade-off for a big savingsaves memory
- Keep a shorter conversationSame model, less to remembershorter
- Load only what's neededFor very large expert-based modelslarge models
- Read from disk as it goesIt still works, just more slowlyslower
$ zmatrix fit ~/Models/Qwen3-8B 8192 chat Rung Verdict Peak As asked will_not_fit 17.87 GiB K/V cache 8-bit will_not_fit 16.23 GiB Stream every block from disk (K/V cache FP32, 8192 tokens) will_not_fit 5.65 GiB Stream every block from disk (K/V cache 8-bit, 8192 tokens) will_not_fit 4.01 GiB Nothing fits, even with the weights streamed from disk. # From the report's notes and suggestions: Predicted peak 17.87 GiB exceeds the 2.23 GiB available. Use 4-bit weights: saves 10.97 GiB, peak 6.90 GiB.
Real output: on a busy Mac with little memory free, ZMatrix explains what won't fit and suggests what will, without loading a thing.
Our promises
Made to be trusted.
Every step of the way.
Your original is always safe
Every change creates a new version, and copies take almost no extra space. Go back, try another idea, or compare any two.
Nothing is a mystery
Every model you make remembers exactly how it was made, so you can repeat it, explain it, or share it with confidence.
Fast because it's native
Built for Apple silicon from the ground up. No web page in disguise and nothing extra to install. A true Mac app.
Private by design
Your models and conversations stay on your Mac. It only goes online to download models, check for updates, or use services you add yourself.
ZMatrix AI Companion for iPhone and Android
Private AI in your pocket.
Even in Airplane Mode.
Download a model made for your phone and chat with it with no signal at all. On iPhone, connect to your Mac with a single scan for something more powerful.
$ zmatrix --version zmatrix 0.0.1 $ zmatrix system-info | grep -E 'chip_name|_cores|memory_pressure"' "performance_cores": 4, "efficiency_cores": 6, "chip_name": "Apple M4", "memory_pressure": "warning", $ zmatrix quantize-model ~/Models/Qwen3-8B ~/Models/Qwen3-8B-Q4 Q4_K_M $ zmatrix serve --model ~/Models/Qwen3-8B-Q4 --port 8080
For developers
Build on it.
Automate it.
The app's tools also run from Terminal, and a built-in server lets your own apps and scripts use your models.
- More than 70 commands for every step, from download to publish
- Works with tools built for OpenAI and Anthropic, pointed at your Mac instead
- Connect your own tools with MCP and Agent Skills
- Automate with Shortcuts, no code needed
Your own AI
starts here.
Download ZMatrix Model Studio and have your first model running in minutes. A friendly setup guide checks your Mac for you.