PrismML Releases Bonsai Image 4B for Local Image Generation

PrismML has introduced Bonsai Image 4B, a family of compressed image generation models designed to run on local devices such as laptops and phones, announced in a press release. The models make high quality diffusion inference practical on hardware ranging from iPhones to Apple Silicon Macs.

Bonsai Image 4B is available in two variants: a 1-bit model and a ternary model. The 1-bit version uses binary transformer weights with group-wise FP16 scaling to achieve maximum compression, shrinking a 4 billion parameter diffusion transformer to 0.93 GB. The ternary variant, which adds a zero state for more representational flexibility, compresses to 1.21 GB. Both retain up to 95 percent of the quality of the full precision model.

On an iPhone 17 Pro Max, Bonsai Image 4B generates a 512 by 512 image in about 9.4 seconds. On a Mac M4 Pro, the same resolution takes about 6 seconds, up to 5.6 times faster than the full precision pipeline. The models are built for local inference across iPhone, Apple Silicon Macs, CUDA GPUs, and small scale serving environments.

Both Bonsai Image 4B variants are released with open weights and code under the Apache 2.0 license. PrismML is also launching Bonsai Studio, an iOS app that allows users to try Bonsai Image 4B directly on Apple Inc. devices.