PrismML released Bonsai Image 4B, a family of compressed image-generation models that run locally on consumer hardware. The system includes both 1-bit and Ternary variants, each using quantized transformer weights to dramatically shrink model size while maintaining quality. This marks a shift toward practical on-device AI generation.

Built from FLUX.2 Klein 4B, the models achieve an 8.3x size reduction for the binary variant (0.93GB) and 6.4x reduction for ternary (1.21GB). These compressed transformers enable generation on devices previously unable to handle full-precision models. The approach keeps the original architecture intact while quantizing weights to {-1, +1} or {-1, 0, +1} formats.

Performance benchmarks show Ternary Bonsai retains 95% of original accuracy across GenEval, HPSv3, and DPG-Bench. The 1-bit version maintains 88% accuracy while staying under 1GB. Generation times clock in at 9.4 seconds for 512x512 images on iPhone 17 Pro Max, with 5.6x faster inference on Mac M4 Pro versus stock pipelines.

Both models release with open weights under Apache 2.0 licensing, accompanied by Bonsai Studio iOS app for direct iPhone testing. This deployment strategy challenges cloud-only generation by enabling private, low-latency iteration without server costs.