A four-bit model is never four bits

Because the scales are never free. In NVFP4, sixteen weights share one eight-bit scale and each picks a four-bit slot on it — 72 bits per sixteen weights, 4.5 effective. Mix in the components a precision map protects at BF16 and the recorded granite artifact lands at 5.65 effective bits per weight. 'This is a 4-bit model' is an incomplete sentence; the complete one is a precision map.

SUPPORTED · recorded — granite-4.1-3b: 4.5 uniform · 5.65 with the late-5 map