ImageOutmake Beta Download Beta

Run local AI models on your Mac.

Download open models for free. Chat, make pictures & videos, code, and speak.

Download for Mac

Free beta. macOS 26 or later, Apple silicon (M1 or later). AI art by the Civitai community · By downloading, you agree to the Terms of Use.

Explore countless AI models.

Before you download anything, each model says how well it runs on your Mac and how fast it writes.

Browse Models

Right here in your browser. No download, no account.

The model store

Ranked for your Mac
Every list puts what runs well on your Mac first. A model too slow or too big for your Mac never ranks first, however popular it is. After that, lists weigh how capable, popular and new each model is.
Fit before you download
Each model says Runs great, Runs well, Tight fit or Too big for your chip and memory, with an estimated speed in words per second.
The right size, picked
Outmake reads each model's GGUF header for its memory and context, then picks the quantization that suits your Mac.
Hand-picked catalog
Qwen3.5, Gemma 4, GPT-OSS, DeepSeek V4, Kimi K2.6, GLM 5.3, MiniMax M2.7, Devstral and more, each file pinned to its checksum.
All of Hugging Face
Search every public model on Hugging Face. What is trending this week shows up in the app with no update.
Downloads that resume
A download survives a quit, a sleep or a dropped connection, and a file is only used once its SHA-256 matches.
Uses what you have
Models you downloaded with LM Studio, Ollama or the Hugging Face cache are used where they are, cloned so they take no extra space.
No setup
Engines install themselves from pinned releases into Outmake's own folder. No Homebrew, no Python, no Terminal.

Outmake Top 10 this week

See all ten
  1. 1Qwen3.8 27BAlibaba Qwen · ★ 4.2
  2. 2Qwen3.8 Flash NextAlibaba Qwen · ★ 4.1
  3. 3Gemma 4 26BGoogle · ★ 4.6
  4. 4Gemma 4 31BGoogle · ★ 4.8
  5. 5MiniMax M3MiniMax · ★ 4.8
  6. 4,200+ moreEach one checked against your Mac
Download and run any model with a click.
Thousands of models and LoRAs to choose from, thanks to Hugging Face and Civitai, with more coming out every day.

Outmake Bench ranks every local model for each Mac: the smartest that run, how fast they write, and what more memory gets you.

Make pictures and videos on your Mac.

Download and use the newest open models and thousands of styles, all for free.

Browse the Gallery

Pictures and video

Eleven model families
FLUX.2 klein, Qwen-Image, Z-Image, Krea 2, FLUX.1, Chroma, SDXL, Illustrious, Pony, Anima and Stable Diffusion 1.5.
Hundreds of models
Browse picture models from Civitai and Hugging Face, with sample pictures from each maker, and download any of them in one click.
Thousands of LoRA styles
Outmake checks that a style applies to your model, remembers the result, and hides a style once it is known not to fit.
Maker settings
Steps, guidance, sampler and scheduler come from what each model's maker publishes.
Inpainting
Paint over part of a picture and say what should be there instead.
MetalFX upscaling
Upscales a picture on the GPU in about 30 milliseconds, with the scaler built into macOS.
Background removal
Vision's foreground mask, the model behind Remove Background in Preview.
Uncensored
Uncensored chat models that answer without the usual refusals, and picture models and styles made for adults. Off until you turn on Mature Content in Settings › Models and confirm you're 18 or older.
Labeled as AI
Every picture carries IPTC's Digital Source Type in its metadata, the label Google and Meta read.
Outmake uses Apple's software to remove backgrounds, and upscales with MetalFX, Apple's on-device upscaler.
Brainstorm, generate and animate scenes, all on your Mac in Outmake.

Wan 2.2 14B made the 2-second wave in 12 minutes on a MacBook Pro with 16 GB of memory. Background removal timed on the same Mac.

Dynamic notch.

A Mac's notch says hello and waves. The pointer moves under it and a small chat opens. Someone types "Move lunch with Sam to Friday and let him know" and sends it. The chat folds back into the notch, which shows each step as it works, then says the reply beside the notch with a tick.

Optimized for Mac via speculative decoding.

When a local model edits code, a small model guesses the next words and the large one checks them all at once. The code comes out the same.

Qwen3 4B editing a code file on a MacBook Pro with M1 Pro and 16 GB of memory. Measured.

Voice control for your Mac.

Chat with Outmake in 100+ languages to operate your Mac, customize its voice and dictate. Voices run on your Mac or online, and online voices cover the most languages.

Play the Voices

Voice control

Computer use
Presses, types and scrolls in any app through macOS Accessibility, including windows behind others. Your pointer stays where you left it.
Screen reading
Reads the window in front of you the way a person does, main content first, so you can ask about what is on it.
Personalized voice
Twelve natural voices built in, and more from Kokoro, Piper and others in the voice store. Change the speed without changing the pitch.
Personalities
Pick how it talks, from Brutally Honest to Rubber Duck. Each one shows its full prompt, and you can write your own.
Head gestures
When it asks a question, nod for yes or shake for no. Works with AirPods that have head tracking.
Press to Mute
Press your AirPods stem once to put it on hold while you talk to someone. Press again to carry on.
Spatial Audio
With head tracking on, its voice stays a meter in front of you as you turn your head.
Barge-in
Talk over it to stop it or change the request halfway through a sentence.
Talk while it works
Ask something quick during a long task. It answers, then carries on with the task.
Turn detection
Waits until you finish the thought, so stopping to think does not cut you off.
Streaming speech
Starts talking as soon as the first sentence of the reply is written.
Wake word
Say its name to start, or press a keyboard shortcut you pick.

Try out these advanced voices from the talented team at Alibaba Qwen.

Dictation in any app.

Hold Right Option and talk. The words are typed where your cursor is, by a speech model on your Mac, or the open-source dictation model of your choice.

Bring your Claude, ChatGPT or Gemini plan.

Use hundreds of cloud models inside Outmake: OpenRouter, DeepSeek, Moonshot, xAI, MiniMax, Xiaomi, Meta, Vercel AI Gateway, OpenCode Zen, Z.ai, Kimi and many more.

Agents

Agent Client Protocol
Claude Code, Codex and OpenCode run inside Outmake, on the plans you pay for.
Model Context Protocol
Add tool servers from the public MCP Registry.
Kernel sandbox
Commands an agent runs are fenced by the macOS sandbox, except in a Gemini chat and at Full Access. A file it is not allowed to touch stays untouched, whatever the agent was told.
Undo any reply
Files a reply changed can be put back, and the undo can be undone.
Schedules
A prompt runs on its own at the time you set, into its own chat, so each run can see the last.
The plan you pay for. Agents run on your own account, and Outmake adds no cost.
Undo anything. Files an agent changed can be put back, and the undo can be undone.

Made for macOS.

Written in Swift, with Liquid Glass. Local models run on your Mac's graphics chip, through Metal and Apple's own frameworks.

What's inside

Engines

Metal acceleration
Chat runs on llama.cpp, voice on MLX and pictures on stable-diffusion.cpp, all on the Apple silicon GPU through Metal.
Speculative decoding
Guesses several tokens ahead and checks them in one pass, using the conversation or the model's own multi-token predictor. The text is identical to what the model would have written.
Flash Attention and 8-bit KV cache
A long conversation takes about half the memory, which leaves room for a bigger model.
Vision
Models with a vision projector read screenshots, photos and scanned pages.
Reasoning models
Models that reason before answering show their reasoning as they write.
Memory management
A model loads when you use it, and the ones in the way unload when a new one needs the room.
Apple Foundation Models
Apple's own on-device model, on Macs with Apple Intelligence, with nothing to download.
File reading
PDFs, Word files, scans, audio and video become text on your Mac, with Apple's on-device speech model for anything spoken.

Several Macs

Distributed inference
Experimental. A model too big for one Mac runs across several joined by Thunderbolt. Each Mac holds a share of the model's layers sized to its memory, and Outmake works out the split. It has run with one Mac lending to itself and not yet across two; it appears after Show All Features in Settings › General.
RDMA over Thunderbolt 5
On Macs with Thunderbolt 5 and macOS 26.2 or later, the model's data moves from one Mac's memory straight into the other's by Remote Direct Memory Access.
Models from your other Macs
Chat with the models on your Mac Studio from your MacBook, over Wi-Fi, Ethernet or Thunderbolt. Your Macs find each other with Bonjour.
Encrypted pairing
Two Macs pair by showing the same six digits. The key is agreed with Curve25519 and never crosses the network, and every connection uses TLS with ChaCha20-Poly1305.
Your accent color. In light, dark, or black.
Immersive full screen. Pictures fill the whole display around the notch, and the chat opens from it.
Metal, MLX and Vision. llama.cpp runs chat on Metal, MLX runs voices on Apple's framework for its own chips, and Vision runs edits.
Distributed inference, experimental. A model too big for one Mac runs across several joined by Thunderbolt, each holding the layers that fit its memory. It's built and not yet tried across two Macs: turn on Show All Features in Settings › General to try it.

Privacy first.

Local models run on your Mac with no network access. Anything that would go to another company asks you first. Read the Privacy Policy

Privacy

On-device inference
Chats and pictures on local models are processed by llama.cpp and stable-diffusion.cpp on your Mac. What you type and generate is not sent anywhere.
Loopback only
The chat engine listens on 127.0.0.1 and requires a random key generated at each launch, so no other device on your network can reach it.
No network access for models
A local model cannot open files, run commands or go online.
No account, no tracking
There is no Outmake account, no advertising, and nothing that identifies you. Usage counts go only while Send Measurements and Errors is on, and setup shows that switch before anything is sent.
Anonymous diagnostics
Speed measurements, download results and crash reports go to Outmake's server only while Send Measurements and Errors is on. Setup shows the switch before anything is sent; it starts on, and off in the EU, the rest of the EEA, the UK and Switzerland, and it is in Settings › Privacy afterwards. Records carry no identifier, internet address, prompt, reply, or name or path of your files, and See What Gets Sent shows the exact records.
Local storage
Chats, pictures and memory are stored in ~/Library/Application Support/Outmake, with file permissions that let only your macOS user read them.
Keychain
API keys are stored in the macOS Keychain on this Mac and never synced.
Per-provider consent
Nothing goes to a cloud company (Anthropic, OpenAI, Google, OpenRouter or any other) until you allow it, per company and per use. Requests go from your Mac to the provider, with no Outmake server in between.
Location removed
GPS coordinates in a photo's metadata are removed before the picture is sent to a cloud model.
On-device wake word
The wake word is recognized on your Mac. Nothing heard before it is kept or sent, and head-tracking angles are used live and never stored.
  • On-device inferencellama.cpp and stable-diffusion.cpp run on this Mac. A local model can't open files, run commands or go online.
  • Loopback onlyThe chat engine listens on 127.0.0.1 and needs a key made at each launch, so nothing else on your network can reach it.
  • Per-company consentBefore anything goes to a cloud company, Outmake asks, once for each company and each use.
  • KeychainAPI keys are stored in the macOS Keychain on this Mac.
  • No account, no trackingThere's no Outmake account, no advertising, and nothing that identifies you. Usage counts go only while one switch is on, and setup shows it before anything is sent.
  • Anonymous diagnosticsSpeed measurements and crash reports go only while Send Measurements and Errors is on. Setup shows the switch first; it starts on, and off in the EU, UK and Switzerland. They carry no identifier and no internet address.

Outmake Beta

Give it a try. You probably won't regret it.

Download for Mac

Free · macOS 26 or later · Apple silicon (M1 or later) · 8 GB of memory or more · AI art credits · By downloading, you agree to the Terms of Use