Baseten published a technical explainer on selecting layer precision for NVFP4 quantization

It describes activation calibration, block scaling over groups of 16 values, architecture-based heuristics, and sensitivity scoring for mixed-precision recipes. Its central warning is that isolated-layer sensitivity scores are not additive: the loss contribution of quantizing a layer depends on what else has already been quantized. The post describes SaturationQuant, using Aumann–Shapley scoring, as a way to incorporate saturation so scores can be summed to estimate a recipe's loss more accurately. This is a vendor-authored technical explanation, not independent validation of comparative performance.

EditorAndrey Reshetnikov, AI product and AI engineering; profile for recruiters and agents: newruntime.com/owner-profile.json