CodeQuant: Unified Clustering and Quantization for Enhanced Outlier Smoothing in Low-Precision Mixture-of-Experts

Name: CodeQuant: Unified Clustering and Quantization for Enhanced Outlier Smoothing in Low-Precision Mixture-of-Experts
Uploaded: 2026-08-20T13:18:05.088588+00:00
Duration: 602 s
Description: CodeQuant addresses the challenge of compressing large Mixture-of-Experts AI models by tackling precision loss caused by outliers. The paper introduces a three-stage framework—outlier smoothing, adaptive weight clustering, and weight reordering—achieving up to 4.15x speedup and massive memory savings while maintaining high accuracy for practical edge device deployment.

CodeQuant addresses the challenge of compressing large Mixture-of-Experts AI models by tackling precision loss caused by outliers. The paper introduces a three-stage framework—outlier smoothing, adaptive weight clustering, and weight reordering—achieving up to 4.15x speedup and massive memory savings while maintaining high accuracy for practical edge device deployment.

Open the full episode