For developers

Your models.
Fully programmable.

What you build by hand in the app, you can script from Terminal with the same engine and the same results. Then serve your models to any app that speaks the OpenAI or Anthropic API.

Get the direct editionLocal server
zsh — a recipe, by hand
# Find a model, download it, and see whether it fits (~/Models/Qwen3-8B stands for where it was saved)
$ zmatrix hub search qwen3 --format safetensors
$ zmatrix hub download Qwen/Qwen3-8B
$ zmatrix fit ~/Models/Qwen3-8B 8192 chat
 
# Shrink it, measure it, then serve it
$ zmatrix quantize-model ~/Models/Qwen3-8B ~/Models/Qwen3-8B-Q4 Q4_K_M
$ zmatrix bench speed --model ~/Models/Qwen3-8B-Q4 --contexts 512,2048
$ zmatrix serve --model ~/Models/Qwen3-8B-Q4 --port 8080 --api-key-env ZM_KEY
70+

Commands, from finding a model to publishing it.

1

Engine. The app and the command line share it, so results always match.

JSON

Output for every job and report, ready for your scripts and CI.

0

Dependencies to install. No Python, no Docker, just the app.

Inspect and fit

system-info · gpu-check · fit · inspect · tensors · detect-model · generate

Quantize and convert

quantize-model · quantize-gguf · quantize-learned · quantize-sensitivity · imatrix · quant-matrix · quant-report · prune · convert · export coreml

Edit

abliterate · abliterate-explore · concept · steer · lora · merge · surgery · tensor-edit · experts · template · versions · diff

Train and measure

tune plan · tune train · bench speed · bench ppl · bench eval · bench needle · viz lens

Library

hub · models · datasets · jobs · storage · recipe · context

Platform

serve · publish · plugins · skills · update · completions · man

The command line comes with the direct download and installs from Settings. The Mac App Store edition doesn't include it.

Local server

Point your apps
at your Mac.

Swap a cloud API for your own Mac by changing one address. Share it on your network with an API key, allow the websites you trust, and keep requests in your own hands.

/v1/chat/completions/v1/messages/v1/embeddings/v1/models/v1/decisions
an example request
$ curl http://127.0.0.1:8080/v1/chat/completions \
    -H "Authorization: Bearer $ZM_KEY" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "Qwen3-8B-Q4",
      "messages": [{"role": "user", "content": "Explain Q4_K_M."}],
      "stream": true
    }'

Extend it

Make it
do more.

MCP

Connect your tools with MCP

Give your models access to the tools you already use. Every tool asks for your approval before it acts, prompts and resources drop straight into your message, and secrets stay in your Keychain. Local stdio servers need the direct download.

Agent Skills

Teach new skills

Import skills in the open Agent Skills format once and use them everywhere. The model picks the right one when it needs it, or you call it by name with a slash.

Plugins

Plugins

Add your own tests, scoring and prompt sets in both editions, and signed native plugins with the direct download.

Automation

Shortcuts

Shrink a model, run a test or start a chat from Shortcuts, no code needed.

Provenance

A record of everything

Every model you export records exactly how it was made, and comes with a ready-to-publish model card for Hugging Face.

Build something
only you could.

Get the direct download for the command line, native plugins and local tool servers.

DownloadSee the app