1-Bit Bonsai Image 4B Image Generation Model for Local Devices

Bonsai Image 4B is a compact image generation model family designed to perform high-quality diffusion inference on local hardware, from laptops to smartphones.

Two variant models are provided:

• 1-bit Bonsai Image 4B: Uses binary weights ({−1, +1}), optimized for environments with heavy memory pressure and bandwidth constraints, with an effective 1.125 bits per weight.
• Ternary Bonsai Image 4B: Uses ternary weights ({−1, 0, +1}), offering higher visual quality and prompt fidelity, with an effective 1.71 bits per weight.

Built on top of FLUX.2 Klein 4B, it dramatically reduces memory usage by changing how weights are represented.

• 1-bit model: The diffusion transformer size is 0.93 GB, about an 8.3x reduction from the original (7.75 GB).
• Ternary model: The diffusion transformer size is 1.21 GB, about a 6.4x reduction from the original.

It supports the MLX path on Apple Silicon (iPhone, iPad, Mac) and the Gemlite kernel on CUDA GPUs, and can run on-device on devices such as the iPhone 17 Pro Max.