PrismML has released Bonsai Image 4B, a family of compact image-generation models designed to run high-quality diffusion inference on local hardware, including laptops and phones.
A Model Built for Local Hardware
Local image generation faces a hard constraint: the model must fit within a device's memory budget. For a 4-billion-parameter image model, the diffusion transformer is the largest component and runs repeatedly during each denoising step. Transformer size directly drives memory pressure, bandwidth demand, and inference speed.
Bonsai Image 4B is built on the FLUX.2 Klein 4B architecture. PrismML kept the architecture intact but changed how the transformer weights are represented. By converting weights into binary and ternary form, the company reduced the part of the image pipeline that matters most for local deployment.
The binary layers provide roughly a 14x reduction relative to full-precision transformer weights. A small set of precision-sensitive supporting tensors, about 5% of the total, remains in FP16. The final 1-bit Bonsai Image 4B transformer is 0.93 GB, an 8.3x reduction from the 7.75 GB full-precision FLUX.2 Klein 4B.
The ternary variant uses the same structure. Its ternary layers deliver about a 10x reduction, and the final Ternary Bonsai Image 4B transformer is 1.21 GB, a 6.4x reduction from the full-precision version. It is slightly larger than the 1-bit model, but the additional zero state improves visual quality and prompt fidelity.
Including the compressed text encoder and FP16 VAE, the Apple Silicon deployment payload is 3.42 GB for 1-bit Bonsai Image 4B and 3.88 GB for Ternary Bonsai Image 4B. The full-precision FLUX.2 Klein 4B requires 15.97 GB. At runtime, the text encoder is offloaded after prompt encoding, so the mean memory usage is smaller than the total payload. For a 512x512 image, mean-active memory is 1.5 GB for the binary model and 1.96 GB for the ternary model, compared to 11.74 GB for the original. For a 1024x1024 image, those figures are 1.95 GB and 2.38 GB, compared to 14.39 GB.
Two Variants for Different Priorities
PrismML offers two variants. The 1-bit model is footprint-oriented, bringing the diffusion transformer below 1 GB. The ternary model is quality-oriented, retaining 95% of the FLUX.2 Klein 4B accuracy across GenEval, HPSv3, and DPG-Bench benchmarks while still reducing the transformer footprint by 6.4x. The 1-bit model retains 88% accuracy.
Together, the two variants shift the quality-footprint frontier. Bonsai Image remains competitive with modern 4B-class image models while using a fraction of their diffusion-transformer footprint, and it substantially outperforms smaller models with similar memory footprints.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Performance and Benchmarks
The deployment stack supports Apple Silicon iPhones, iPads, and Macs as well as CUDA GPUs, using MLX low-bit paths on Apple hardware and Gemlite low-bit GEMM kernels on CUDA. On an iPhone 17 Pro Max, the full-precision FLUX.2 Klein 4B pipeline does not fit within the device memory budget, but both Bonsai Image variants run on-device. A 512x512 image generates in 9.4 seconds on that iPhone and about 6 seconds on a Mac M4 Pro. On the Mac M4 Pro, Bonsai Image 4B is up to 5.6x faster than the stock full-precision MFLUX pipeline.
A Shift in Deployment Paradigm
Image generation is not only a model-quality problem but also a deployment problem. Cloud APIs remain the right choice for many products, but cloud-only generation imposes constraints: every prompt is a remote request, each iteration carries marginal serving cost, and every interaction adds round-trip latency.
Local inference changes that dynamic. Once the model fits on the device, generation can sit directly inside the product experience. It becomes cheaper to run, faster to iterate on, and easier to use in environments where prompts and generated assets should remain private.
Bonsai Image 4B is a step toward that deployment regime: capable image generation running closer to the user, on hardware they already own.
Availability and Open Weights
Both 1-bit and Ternary Bonsai Image 4B will be released with open weights and code under the Apache 2.0 license. PrismML is also launching Bonsai Studio, an iOS app for trying Bonsai Image 4B directly on an iPhone.
PrismML emerged from a team of Caltech researchers and was founded with support from Khosla Ventures, Cerberus, and Google. The company has spent years working on compressing neural networks without sacrificing reasoning ability.