PrismML releases 1-bit and ternary Bonsai Image 4B for local image generation

PrismML has released Bonsai Image 4B, a family of compact image-generation models built to run diffusion inference on local hardware, from laptops down to phones. The model ships in two variants: 1-bit Bonsai Image 4B, which uses binary {−1, +1} transformer weights with an FP16 group-wise scaling factor, and Ternary Bonsai Image 4B, which adds a zero state with {−1, 0, +1} weights.

What the two variants change

The binary variant reaches 1.125 effective bits per weight and targets maximum compression, where memory pressure, bandwidth and deployment footprint are the main constraints. The ternary variant lands at 1.71 effective bits per weight; the extra zero state gives it more representational flexibility, improving visual quality and prompt fidelity while staying extremely compact. Both models are built from FLUX.2 Klein 4B: the architecture is unchanged, but the transformer weights are represented in binary or ternary form.

Footprint and memory

The diffusion transformer is the largest part of a 4B-class image model and runs repeatedly during generation, since every denoising step invokes it again. In full precision, the FLUX.2 Klein 4B transformer takes 7.75 GB. Bonsai cuts that to 0.93 GB for the 1-bit model — an 8.3x reduction — and 1.21 GB for the ternary model, a 6.4x reduction. A small set of precision-sensitive projection layers, about 5% of the tensors, stays in FP16. Including the compressed text encoder and an FP16 VAE, the Apple Silicon deployment payload is 3.42 GB and 3.88 GB, against 15.97 GB for full precision.

Mean-active memory drops accordingly. For a 512x512 image the binary and ternary models use 1.5 GB and 1.96 GB, compared with 11.74 GB for the original model; at 1024x1024 the figures are 1.95 GB and 2.38 GB against 14.39 GB. The deployment stack covers Apple Silicon iPhones, iPads and Macs through MLX low-bit paths, plus CUDA GPUs via Gemlite low-bit GEMM kernels.

Speed and benchmarks

On an iPhone 17 Pro Max, where the full-precision pipeline does not fit within the device memory budget, Bonsai generates a 512x512 image in 9.4 seconds; on a Mac M4 Pro it takes about 6 seconds and runs up to 5.6x faster than the stock full-precision MFLUX pipeline. PrismML evaluated the models on GenEval, HPSv3 and DPG-Bench: the ternary variant retains 95% of FLUX.2 Klein 4B accuracy while cutting the transformer footprint 6.4x, and the 1-bit variant retains 88% with the transformer below 1 GB.

Availability

Both variants will be released with open weights and code under the Apache 2.0 license, alongside Bonsai Studio, an iOS app for running the model on an iPhone. PrismML says Bonsai Image 4B is the first image model in its parameter class to run directly on an iPhone. The company emerged from a team of Caltech researchers and was founded with support from Khosla Ventures, Cerberus and Google.