PrismML's Bonsai 4B Packs Image Generation Into a 0.93GB Model

PrismML has open-sourced Bonsai Image 4B, an image generation model built to run directly on edge devices like laptops and smartphones. It comes in two aggressively compressed variants: a 1-bit version and a ternary version, with core diffusion transformers weighing just 0.93GB and 1.21GB, respectively. On Apple Silicon, the full deployment package takes up 3.42GB and 3.88GB. The company is also launching an iOS app, Bonsai Studio, for fully offline image generation on iPhones — no subscription, no API calls to the cloud.

The 1-bit variant targets low-memory hardware, shrinking the diffusion transformer — originally based on the 7.75GB FLUX.2 Klein 4B model — to just 0.93GB, an 8.3x reduction over full precision. The ternary version, at 1.21GB, is 6.4x smaller than full precision but uses a three-valued quantization scheme (-1, 0, +1) for more expressive flexibility, improving image quality and prompt fidelity while keeping storage low.

Despite the extreme compression, PrismML says Bonsai Image 4B holds its own against much larger mainstream image generators in object composition, human preference, aesthetic quality, and complex prompt following. The 1-bit and ternary versions retain 88% and 95% of the full-precision model's performance, respectively. Both variants are released under the Apache 2.0 open-source license.