Early-Signal Detector

method · first seen 4 Oct 2026

post-training quantization

Also: AWQ, PTQ, NF4 quantization, full integer quantization

Compressing a trained neural network to lower bit-widths without retraining or access to training data.

Post-training quantization compresses an already trained network to lower bit-widths (int8, int4, FP4 and below) using at most a small calibration set, with no retraining. The sources are a steady stream of incremental arXiv work rather than a new idea: a framework for looped language models, a study of weight averaging during pretraining and its effect on PTQ, Huawei's MACaQ (an MXFP4 W4A4 method released with code), work on calibration data for sub-4-bit models, measurement-noise studies, and edge deployment papers built on GPTQ and AWQ formats. Applications now extend to vision-language-action policies and robot world models.

PTQ is already a standard step for deploying large models. The newer work matters mainly for 4-bit-and-below weight-activation formats such as MXFP4 and for checking whether quantized models keep their robustness, not just their headline accuracy.

Analyst note (established), 17 Nov 2026.

Mentions per week

01819 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 126 on 14 Nov 2026 · 46 mentions in 28 days · 46 independent groups · papers → code · first seen 6 weeks ago · acceleration +0.24

Earliest sources

Latest sources

Mentioned together with

GPTQvision-language-action modelmixed-precision quantizationrotation-based quantizationself-hosted LLM deploymentactivation-aware weight quantizationquantization-aware trainingspeculative decodingvision-language modellarge language model quantization