PrismML unveils 1-bit and ternary Bonsai Image 4B models for local image generation on laptops and phones

PrismML released Bonsai Image 4B, a compact family of image-generation models designed for local diffusion inference on devices ranging from laptops to phones. The models use 1-bit and ternary transformer weights to reduce memory and bandwidth needs while preserving usable image quality, with open weights and claimed iPhone support.

The release points to a broader shift toward on-device generative AI, where compression techniques can make previously cloud-bound workloads practical on consumer hardware. If quality and speed hold up, models like Bonsai could reduce reliance on hosted inference and expand privacy-preserving creative tools at the edge.