Post-Training Quantization

TL;DR for operators The useful surprise in Tail-Aware HiFloat4 is not that a 4-bit video model gets worse. That part is not exactly a Nobel-level plot twist. The useful surprise is where it gets worse. The paper reports a W4A4 HiFloat4 post-training quantization pipeline for Wan2.2-I2V-A14B, and under matched generation settings the unweighted mean score drops from 0.6800 to 0.5880. But the collapse is concentrated: subject consistency falls from 0.9331 to 0.5324, while aesthetic quality is effectively unchanged, overall consistency is comparable, and motion smoothness drops only slightly from 0.9923 to 0.9803.1 ...

Compression sounds simple until the model starts forgetting how to think. A deployment team takes a large language model, squeezes its weights into lower precision, saves memory, improves serving economics, and expects the model to behave like a slightly thinner version of itself. Then INT4 arrives with a polite smile and removes just enough reasoning ability to make the business case awkward. The model still answers. It still looks fluent. It just becomes less reliable exactly where the product needed it to stay sharp. ...