PrismML releases Bonsai Image 4B, for local image generation on iPhone and Mac

PrismML releases Bonsai Image 4B, a compressed FLUX.2 Klein model running locally on Mac M4 Pro and iPhone 17 Pro Max with up to 95% of original quality.

PrismML releases Bonsai Image 4B, a family of image generation models designed to run directly on personal devices, such as laptops and phones. The company, stemming from Caltech researchers, has specialized in model compression, having already applied this approach to language models.

Bonsai Image adopts the architecture of FLUX.2 Klein 4B, developed by Black Forest Labs, but lightens its heaviest component, the diffusion transformer used at each generation step. Two variants coexist. The 1-bit version, aimed at maximum compression, reduces this transformer to 0.93 GB, which is 8.3 times less than the full-precision version's 7.75 GB. The ternary version, which adds a neutral state to the weights, weighs 1.21 GB for better rendering and prompt fidelity. By integrating the text encoder and decoder, the entire package occupies approximately 3.4 to 3.9 GB, compared to nearly 16 GB for the original model.

This slimming down opens up previously inaccessible territories: according to PrismML, Bonsai Image would be the first model in its category to run directly on iPhone. A 512×512 image takes approximately 6 seconds on a Mac M4 Pro and 9.4 seconds on an iPhone 17 Pro Max. On the GenEval, HPSv3, and DPG-Bench benchmarks, the ternary variant retains 95% of the full model's quality, the 1-bit version 88%.

The weights and code are released open source under the Apache 2.0 license, accompanied by an iOS application, Bonsai Studio, for testing generation on iPhone.