Prism ML Bonsai 4B: Ternary Image Model Runs at 3.7 GB

Prism ML Bonsai 4B: Ternary Image Model Runs at 3.7 GB

Prism ML has released Bonsai Image 4B, a binary/ternary retrain of Black Forest Labs' Flux 2 Klein 4B. Unlike quantization approaches that collapse quality, Prism rebuilds the diffusion transformer weights natively for binary/ternary representation. The ternary variant peaks at approximately 3.7 GB during generation — down from ~13 GB for FP16 — producing usable images in under 5 seconds on a MacBook at roughly 95% quality. Text rendering remains a known weak spot.

Why It Matters

A high-quality image generation model running in under 4 GB of RAM is the threshold that makes local image gen viable for consumer hardware without dedicated GPU setups. Bonsai Image 4B also runs on iOS via Bonsai Studio, pushing local multimodal generation to phones.