Bonsai Image 4B: The Free Open-Source AI Image Generator That Runs Offline on Your iPhone

For the past few years, generating an image with AI has meant staying online. You open a website or an app, type a prompt, and your request travels to a server farm somewhere in the cloud before the picture comes back. Every image you make passes through a machine you don't own.

A new open-source model moves that entire pipeline onto the phone in your pocket.

An on-device image generator that skips the cloud entirely

The model is called Bonsai Image 4B, and it was built by PrismML, a startup spun out of Caltech. It runs end to end on iPhone, iPad, and Mac — completely offline, with nothing uploaded to a remote server.

That's a harder trick than it sounds. Image-generation models are typically huge, and a phone can neither store nor run them. PrismML's breakthrough was putting the model on a strict diet.

How they shrank a 4-billion-parameter model to fit a phone

Bonsai Image 4B is built on FLUX.2 Klein, a 4-billion-parameter open-source image model. In its original form it needs nearly 12GB of memory to run — more than almost any phone can spare.

PrismML's technique is called low-bit quantization. In plain English: the original model stored every number in fine-grained detail, and the team simplified those numbers down to little more than "positive" or "negative" states. The model shrank more than eightfold, dropping to roughly 1.5GB at runtime — small enough for a recent flagship phone to handle.

The trade-off is image quality. The most aggressively compressed build produces images at about 90% of the original's fidelity; a version that keeps a little more precision lands around 95%. For anyone generating images casually, that gap is essentially invisible.

Why offline AI image generation matters

Running an image generator locally, with no internet connection, delivers three concrete advantages.

It's free. The model is released open-source under the Apache 2.0 license, so anyone can download and use it. PrismML also ships an iPhone app, Bonsai Studio: load the model file, type a description, and you get an image.

It works with no signal. On a flight, in a dead zone, or any time you'd rather not upload your work, it keeps generating. Your prompts and your finished images never leave your device — a real win for privacy.

It's fast enough. Expect roughly nine seconds per image on an iPhone 17 Pro Max, and about six seconds on an M4 Mac.

The honest limitations

This isn't magic. The model file eats up three to four gigabytes of storage, so the initial download takes patience. Generation is still a step slower than a top-tier cloud service, and the default output resolution is on the small side.

But the genuinely interesting part is what it proves: a model that once demanded data-center GPUs now fits inside a phone — for free.

For two years, the trend in AI has run in one direction: bigger, pricier, more dependent on the cloud. Bonsai Image 4B runs the opposite way.

Once image generation, writing, and translation can all ride offline in your pocket, AI finally lands in ordinary people's hands.