PrismML open-sources the 0.93GB image generation model Bonsai4B for on-device use.

ME AI News: According to monitoring by Beating, PrismML’s open-source on-device image generation model, Bonsai Image 4B, enables direct diffusion inference on local devices such as laptops and smartphones. Bonsai Image 4B includes two ultra-compact versions: 1-bit and Ternary, with core diffusion Transformer sizes of only 0.93 GB and 1.21 GB, respectively. The complete deployment packages on Apple Silicon devices are 3.42 GB and 3.88 GB, accompanied by the iOS app Bonsai Studio for offline on-device image generation. The 1-bit version is specifically designed for low-memory hardware, reducing the diffusion Transformer size—from the original 7.75 GB FLUX.2 Klein 4B—to just 0.93 GB, achieving an 8.3x reduction compared to the full-precision version. The Ternary version’s core Transformer is 1.21 GB, a 6.4x reduction from full precision, while introducing ternary weights {-1, 0, +1} during quantization to enhance expressive flexibility, improving image quality and prompt fidelity while maintaining low storage requirements. Evaluations show that despite significant size reduction, Bonsai Image 4B still matches mainstream image generation models with far larger parameter counts in object composition, human preference, aesthetic metrics, and complex prompt adherence. The 1-bit and Ternary versions retain 88% and 95% of the full-precision model’s performance, respectively. The accompanying mobile app, Bonsai Studio, allows users to generate images entirely offline on iPhone, with all processing occurring locally—no subscription required and no API requests sent to the cloud. PrismML has now open-sourced both the 1-bit and Ternary versions of the model under the Apache 2.0 license. (Source: BlockBeats)