PrismML Bonsai Image 4B goes browser-local

This looks like a serious push to make image generation practical on consumer hardware, not just a smaller checkpoint with a marketing hook.

• –Binary and ternary weights are the real story here: the model is optimized for memory footprint and local inference, not just raw leaderboard chasing
• –The browser WebGPU demo matters because it lowers the barrier from “interesting model” to “actually usable on-device experience”
• –A ~3GB image model is small enough to change the deployment conversation for laptops, edge devices, and privacy-sensitive workflows
• –Apache-2.0 makes it easier for builders to ship around it without the usual open-weights licensing friction
• –The open question is quality-vs-efficiency tradeoff versus larger image models like FLUX-class systems; if the output is good enough, the packaging wins