PrismML Releases 1-Bit and Ternary Bonsai Image 4B Models
Original: 1-Bit Bonsai Image 4B Image Generation for Local Devices
Why This Matters
Enables high-quality image generation on mobile devices previously impossible
PrismML released Bonsai Image 4B, compact image generation models designed for local devices. The 1-bit variant uses binary weights with 8.3x memory reduction, while ternary version offers 6.4x reduction with better quality.
PrismML launched Bonsai Image 4B family, featuring 1-bit and ternary variants for local image generation on devices from laptops to phones. Built from FLUX.2 Klein 4B architecture, the 1-bit model uses binary {-1, +1} transformer weights achieving 0.93 GB footprint (8.3x reduction from 7.75 GB original). The ternary variant uses {-1, 0, +1} weights at 1.21 GB (6.4x reduction) with improved visual quality. Both models maintain FP16 scaling factors and projection layers for precision-sensitive operations. Total deployment payload is 3.42 GB for 1-bit and 3.88 GB for ternary versions versus 15.97 GB for full precision. Runtime memory usage for 512x512 generation drops to 1.5 GB and 1.96 GB respectively, compared to 11.74 GB original. The company claims this is the first 4B-parameter image model capable of running directly on iPhone.