Beta For Macs with Apple silicon

Run local models. See every turn.

ModelDeck is a Mac app that runs the AI models on your Mac with llama.cpp, MLX and Splash, and shows you what they are doing. Check memory fit before you load, see the exact engine command, and watch time to first token and tokens per second on every reply.

Watch the film Beta status

ModelDeck is beta software provided “as is,” without warranties, to the extent permitted by applicable law. Use it and any connected AI models at your own risk, and verify outputs before relying on them. Read the full notice.

Beta 0.8.0-beta.1 Apple silicon macOS 13 or later Download opens after Apple notarization

Engines install separately: llama.cpp and Splash from Homebrew, MLX with a setup you approve. What to install

ModelDeck chat screen streaming a reply from a local Qwen model. A bar under the reply shows the live state: decoding, tokens so far, tokens per second and time to first token. — Real screenshot: a reply streaming, with the live readout under the chat. Measured once on an M5 Max (64 GB), Qwen3.8-2B Q8_0 GGUF, 2026-10-17. Your numbers will differ.

The film

A short tour

From the models already on your Mac to a live reply and the local server, in 1:12. English captions are available.

ModelDeck launch film, 1:12, with English captions.

What it does

One app for the models on your Mac

Every screenshot below is the real app, cropped to the part that matters.

Your models

Everything already on your Mac, in one list

ModelDeck finds the models already on your Mac and runs each one with its engine: llama.cpp for GGUF, MLX for MLX models, and Splash for Splash models.

GGUF
MLX
Splash

The Models screen listing seven models found on this Mac, each with its engine (GGUF, Splash or MLX), quantization, size and context length, and a Load button.

Before you load

Know it fits before you load it

The load sheet estimates the memory a model needs with your context size, counts what other apps already use, and tells you whether it fits.

Settings are in plain words, each with the engine flag beside it. The exact command line is one click away.

The load sheet for a GGUF model: a memory bar marked Fits, a breakdown of memory in use, model, context and runtime, and plain-language settings such as context size and GPU layers with their engine flags.

While it runs

See what happened on every turn

Time to first token and tokens per second update live while a reply streams, and stay with the reply when it ends. The inspector shows the exact request, the raw stream, the response and the timing behind any reply.

Reasoning, tool calls, edit and regenerate are all there when you need them.

A finished chat reply with its timing line, next to the inspector panel showing the exact JSON request sent to the engine. — Measured once on an M5 Max (64 GB), Qwen3.8-2B Q8_0 GGUF, 2026-10-17. Your numbers will differ.

For your tools

One local endpoint for your tools

Point any OpenAI-compatible client at ModelDeck. It listens on this Mac only and logs each request with its timing.

It can load a model on the first request and unload it when idle, and an API key is optional. Close the window and the server keeps answering from the menu bar.

The Server screen: running on a loopback address, configuration for port, optional API key and idle unload, served models, a ready-to-copy curl example, and a live request log with time to first token and tokens per second. — Timings in the request log come from one capture run on an M5 Max (64 GB). Your numbers will differ.

Find the next model

Search Hugging Face, with fit shown for your Mac

Browse and search Hugging Face from inside the app, see which models fit in memory, and download with resume and a checksum check.

The Discover screen listing Hugging Face models with family, parameters, quantization, size, license, downloads and a fit badge reading Fits or Won't fit.

And the rest

Benchmarks on your Mac

Run quick, standard or long-prompt benchmarks on a model file, on your own hardware.
Evaluate

Score yes or no, choice and rating tasks, and run test suites against a model.
Keyboard first

⌘K for commands. ⌘L to load a model, with fit shown for each one.
No telemetry

ModelDeck sends no telemetry. Settings says so, next to the version you are running.

One measured run

Time to first token: 48 ms
Decode speed: 166 tokens/s
Full reply: 210 tokens in 1.3 s

Measured once on an M5 Max (64 GB), Qwen3.8-2B Q8_0 GGUF, 2026-10-17. Your numbers will differ.

Beta

ModelDeck beta

For Macs with Apple silicon. Engines install separately.

ModelDeck 0.8.0-beta.1 Beta

The download is not available yet. We publish the disk image here once Apple has notarized it. This page will show its size and SHA-256 at that point.

Version: 0.8.0-beta.1 (1)
Channel: Beta
Platform: Apple silicon (arm64)
Requires: macOS 13 or later
Signing: Waiting on Apple Developer ID signing and notarization

Engines install separately

ModelDeck does not include engines or models. Install the engines for the models you want to run.

GGUF: brew install llama.cpp
Splash: brew install incoai/tap/splash
MLX: needs Python 3.10 or newer, for example brew install python. The first time you load an MLX model, ModelDeck shows the setup commands and runs them only if you agree. It creates a private Python environment and downloads mlx-lm from the Python Package Index.

Good to know

This is beta software. Expect rough edges, and verify model outputs before relying on them.
No automatic updates. Check this site for new versions.
Apple silicon only. There is no Intel, Windows or Linux version.
The macOS 13 floor comes from the app framework. Testing so far has been on macOS 27.

Release notes for 0.8.0-beta.1

ModelDeck is a Mac app for running local AI models and seeing what they do on every turn. This is a beta release for Macs with Apple silicon.

In this beta

Finds the GGUF, MLX and Splash models already on your Mac, and runs them with llama.cpp, MLX or Splash.
Tells you whether a model fits in memory before you load it, counting memory that other apps use.
Plain load settings, with the exact engine command one click away.
Streaming chat with time to first token and tokens per second on every turn, plus reasoning, tool calls, edit and regenerate.
An inspector with the exact request, stream, response and timing behind any reply.
A local OpenAI-compatible endpoint for your tools, on this Mac only, with a per-request log, an optional API key, loading on first request and unloading when idle.
Hugging Face search, with resumable downloads and a checksum check.
Benchmarks on your own Mac, plus a decision lab and test suites for scoring models.
⌘K for commands and ⌘L to load a model.
Close the window and the server keeps answering from the menu bar.
ModelDeck sends no telemetry.

Engines are installed separately

ModelDeck does not include model engines or models. Install the engines for the models you want to run:

GGUF models: install llama.cpp with Homebrew (brew install llama.cpp). ModelDeck uses its llama-server.
Splash models: install Splash with Homebrew (brew install incoai/tap/splash).
MLX models: ModelDeck offers to set up a private Python environment in ~/.modeldeck/mlx-venv the first time you load an MLX model. It shows the commands first and runs them only if you agree. It needs Python 3.10 or newer for Apple silicon, for example brew install python or the macOS installer from python.org. ModelDeck finds it but never installs Python itself. The setup downloads mlx-lm 0.32 from the Python Package Index (pypi.org), about 300 MB plus dependencies.

Requirements

A Mac with Apple silicon. Intel Macs are not supported.
macOS 13 Ventura or later. This floor comes from the app framework. Testing so far has been on macOS 27.
Disk space and memory for the models you choose. ModelDeck shows whether a model fits before you load it.

Install

Open the downloaded disk image.
Drag ModelDeck to Applications.
Open ModelDeck from Applications.

To check the download, run shasum -a 256 on the disk image and compare the result with the SHA-256 on the download page or in the .sha256 file next to the download.

Known limitations

This is beta software. Expect rough edges, and verify model outputs before relying on them.
No automatic updates. Check the ModelDeck site for new versions.
Apple silicon only.
Engines and models are installed separately, as described above.

Network use

ModelDeck sends no telemetry. Discover connects to Hugging Face when you browse, search or download models, and the MLX setup, which runs only after you approve it, downloads mlx-lm from the Python Package Index (pypi.org). Models run on your Mac.

The install steps in these notes apply once the disk image is published here.

FAQ

Questions, answered plainly

Does ModelDeck send telemetry?

No. ModelDeck sends no telemetry. Discover connects to Hugging Face when you browse, search or download models, and the MLX setup, which runs only after you approve it, downloads mlx-lm from the Python Package Index (pypi.org).

Where do the models run?

On your Mac. The engines and the local endpoint listen on this Mac only.

Does it include the engines?

No. Install llama.cpp and Splash with Homebrew. For MLX, ModelDeck offers a one-time setup that creates a private Python environment, and runs it only if you agree. It needs Python 3.10 or newer. See the commands.

Which Macs does it run on?

Macs with Apple silicon on macOS 13 Ventura or later. There is no Intel, Windows or Linux version.

Does it update itself?

No. This beta has no automatic updates.

Is it finished?

Not yet. ModelDeck is in beta, so expect rough edges, and verify model outputs before relying on them. Read the beta notice.