Prism ML has officially released Bonsai Image 4B, shrinking a 4-billion-parameter diffusion transformer down to 0.93 GB using 1-bit quantization. The open-weight model family enables local image generation on devices like iPhones and Macs without relying on server-side APIs. Requiring as little as 1.5 GB of active memory, the 1-bit variant produces 512x512 images in 9.4 seconds directly on an iPhone 17 Pro Max.
Extreme Compression Mechanics
Prism ML, which emerged from Caltech research, built Bonsai Image 4B using the FLUX.2 Klein 4B architecture as its foundation. While maintaining the core transformer design, the team altered weight representation to drastically reduce memory overhead during iterative denoising steps.
The family offers two main variants for local execution:
- 1-bit Bonsai Image 4B: Converts transformer weights to binary with FP16 group-wise scaling (1.125 effective bits per weight), cutting the transformer footprint by 8.3x to 0.93 GB while retaining 88% of full-precision performance.
- Ternary Bonsai Image 4B: Employs ternary weights with FP16 group-wise scaling (1.71 effective bits per weight), shrinking the transformer to 1.21 GB while preserving 95% of full-precision output quality.
- Projection Layer Preservation: Approximately 5% of precision-sensitive supporting tensors remain in FP16 to stabilize visual outputs and prevent severe degradation.
- Reduced Active Memory: On-device active memory requirements for 512x512 generation drop to 1.5 GB for the binary model and 1.96 GB for the ternary model, down from 11.74 GB in the original FP16 pipeline.
On-Device Performance and Benchmarks
By targeting memory pressure, Bonsai Image 4B shifts the deployment boundaries for generative media. On Apple Silicon, total deployment payloads including the compressed text encoder and FP16 VAE sit at 3.42 GB for the binary model and 3.88 GB for the ternary model, compared to 15.97 GB for full-precision FLUX.2 Klein 4B.
Benchmarks evaluated across GenEval, HPSv3, and DPG-Bench show 1-bit Bonsai Image 4B retains 88% of full-precision FLUX.2 Klein 4B performance while reducing diffusion transformer memory footprint by 8.3x. On a Mac M4 Pro, generation finishes in roughly 6 seconds, representing a 5.6x speed improvement over stock full-precision MFLUX execution.
Implementation Steps and Operational Caveats
Developers looking to deploy Bonsai Image 4B can access open weights and code licensed under Apache 2.0. The implementation leverages Apple MLX low-bit execution paths on Apple Silicon and Gemlite low-bit GEMM kernels on NVIDIA CUDA GPUs. Prism ML has also launched an iOS application called Bonsai Studio for immediate testing on mobile hardware.