MIT and NVIDIA QDA agent automates quantization recipe search
The hidden cost in quantization has always been evaluation, and shifting it to minutes-scale KL-divergence scoring changes who can run systematic search from only large labs to teams with modest GPU budgets. Practitioners deploying reasoning models at low precision get a replicable recipe-finding agent rather than hand tuning. The GLM-5.3-Flash result at NVFP4 W4A4 is directly comparable to vendor checkpoints, useful for anyone choosing between published quants and custom ones.