T2Quant enables task-aware mixed-precision quantization for Diffusion Language Models
October 21, 2026
T2Quant uses a label-free sensitivity signal to allocate activation precision across denoising steps in Diffusion Language Models (DLMs). This method addresses the time-dependent quantization sensitivity of DLMs to improve accuracy under fixed bit-width budgets.
HOW THIS AFFECTS YOU
●
builderYou can reduce the memory footprint of DLMs without significant accuracy loss by applying time-aware precision.
●
researcherThis framework provides a method for predicting quantization sensitivity without requiring ground-truth labels.