Just saw PrismML did extreme quantization on FLUX.2 Klein 4B, 1-bit can squeeze it down to 0.93GB, ternary version is 1.21GB, Apache 2.0 license. They’re claiming it’s the first 4B-level image model that can run on a phone. What I care about is the trade-offs behind the numbers: the original FP16 transformer needs 7.
75GB, around 13GB VRAM, a 4070 can run it but phones are a no-go; after quantization, the size gets cut to one-sixth. Quality-wise, they admit it inherits the ceiling of the Klein distilled version, and 1-bit shows visible degradation to the naked eye. So my take is—if you’ve got a dedicated GPU on desktop, go for the ternary version, 90% quality at one-sixth the size, best bang for your buck; 1-bit only makes sense on phones or devices with under 2GB VRAM.
This is a win for portability, not a breakthrough in image quality, don’t expect it to look better than the original.