PrismML introduces Bonsai Image 4B, a 4B-parameter diffusion image generation model optimized for local inference on devices like iPhones, iPads, Macs, and CUDA GPUs.

It comes in two low-bit variants: a 1-bit model with binary transformer weights that reduces the diffusion transformer footprint to 0.93 GB (8.3x smaller than FLUX.2 Klein 4B) and a ternary model at 1.21 GB (6.4x smaller) that preserves about 95% of the original model’s benchmark performance. These models enable on-device generation of 512x512 images in around 9.4 seconds on iPhone 17 Pro Max and about 6 seconds on Mac M4 Pro while substantially lowering memory usage and maintaining competitive image quality versus other 4B-class and smaller models.