Prism ML launches Bonsai Image 4B variants

This is a deployment-first release, not just a quantization demo. The notable shift is making a 4B-class image model small enough to run on consumer devices without the cloud.

• –The binary variant is the headline: it pushes the diffusion transformer under 1 GB, which is the kind of footprint that changes where image generation can run.
• –The ternary variant looks like the practical default if quality matters more than absolute minimum memory use.
• –The strongest product angle is local, private iteration on-device, especially for laptops and phones where round trips and API costs slow the creative loop.
• –The iPhone support is a meaningful proof point, but the real test will be whether output quality and UX hold up across real workflows, not benchmark tables alone.