1-bit Bonsai Image 4B Image Generation for Local Devices

Bonsai Image 4B is a family of compact models designed to perform high-quality image generation on local hardware such as laptops and smartphones. Built on the FLUX.2 Klein 4B architecture, it quantizes the diffusion transformer weights to 1-bit or ternary precision, dramatically reducing memory usage.

Memory and Model Size

• 1-bit model: Reduces transformer size to 0.93GB (an 8.3x reduction versus the original).
• Ternary model: Reduces transformer size to 1.21GB (a 6.4x reduction versus the original).
• Runtime memory: When generating 512×512 images, the 1-bit model uses only 1.5GB and the ternary model only 1.96GB of average active memory, enabling smooth execution even on mobile devices.

Performance and Hardware Support

• Inference speed: Generates a 512×512 image in about 6 seconds on a Mac M4 Pro, up to 5.6x faster than MFLUX. On an iPhone 17 Pro Max, it takes 9.4 seconds.
• Quality retention: The ternary model retains 95% of the performance of FLUX.2 Klein 4B, and outperforms SDXL and Stable Diffusion 1.5 on major benchmarks such as GenEval and HPSv3.
• Technology stack: Supports the MLX low-bit path on Apple Silicon and the Gemlite low-bit GEMM kernel on CUDA.