You can now make an AI image on a high-end iPhone without sending anything to the internet. A model called Bonsai Image 4B runs the whole job on the phone, turning your typed words into a 512-by-512-pixel picture in about 9.4 seconds on an iPhone 17 Pro Max. PrismML released it on May 26, 2026, alongside an iPhone app called Bonsai Studio for trying it on the device.
Shrinking it without wrecking the pictures
The hard part wasn't making the model small. It was making it small without wrecking the pictures. An AI model is a giant pile of numbers, called weights, that hold everything it learned, and Bonsai stores most of its 4 billion of them far more crudely than normal to shrink the file.
PrismML ran it through three standard tests that score how good the output looks and how closely it follows the prompt. Its higher-quality version scored about 95 percent as well as the full-size model it is based on, and its smaller version about 88 percent. So the quality drop is modest, and the heaviest part of the model shrinks six to eight times.
Bonsai is a stripped-down version of an open model called FLUX.2 Klein 4B. PrismML changed only how the weights are stored, not the design. The heaviest part, the engine that draws the picture, drops from 7.75 GB to 0.93 GB. The full original cannot fit in an iPhone 17 Pro Max's memory, while Bonsai can, and that is what lets it run on a phone.
Running locally is the real payoff. Image making is a loop where you try a prompt, tweak it, try again, and throw most of the results away. In the cloud every try is a request to a server that costs money and makes you wait, while on the phone that loop is free and private.
PrismML says it will release the weights and code under the open Apache 2.0 license, so others could build their own apps on top of them. The catch is real though. The phone demo's pictures are 512 by 512 pixels, small by today's standards, and about 9.4 seconds each is patient work, not instant.
