How accurate is the model?
#3
by dw1932 - opened
I would like to know how the accuracy of the model is after being quantified. Has anyone conducted a test?
Yes — on held-out WikiText, full-vocab KL-divergence vs the bf16 master ≈ 0.151. For reference, a solid int4 W4A16 (my own quant https://huggingface.co/spectator2026/MiniMax-M3-AWQ-int4) of the same model is ~0.16 on that metric, so this NVFP4 is a touch better. (Measured weight-only on A100).