PrismML launches Bonsai Image 4B, an image model that fits in 0.93GB

PrismML just dropped the Bonsai Image 4B series, a set of image generation models built for local devices. Based on FLUX.2 Klein 4B, they shrink the model size by changing how weights are stored. The 1-bit version uses binary weights and takes only 0.93GB, which is about one-eighth of the original. The Ternary version uses three-value weights, keeps 95% of the accuracy, and fits in 1.21GB.

For speed, it generates a 512x512 image in about 9.4 seconds on an iPhone 17 Pro Max, and around 6 seconds on a Mac M4 Pro. That's up to 5.6 times faster than the full-precision version. Memory usage is also way lower: generating a 1024x1024 image needs just 1.95GB to 2.38GB, compared to 14.39GB for the original model.