GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs

arXiv cs.LG Papers

Summary

Proposes GEMQ, a global expert-level mixed-precision quantization method for MoE LLMs that uses linear programming and router fine-tuning to reduce memory and accelerate inference with minimal accuracy degradation.

arXiv:2605.23078v1 Announce Type: new Abstract: Mixture-of-Experts Large Language Models (MoE-LLMs) achieve strong performance but incur substantial memory overhead due to massive expert parameters. Mixed-precision quantization mitigates this cost by allocating expert-wise bit-widths based on their importance, approaching the accuracy-memory Pareto frontier and enabling extreme low-bit quantization. However, existing methods rely on layer-wise importance estimation and overlook router shifts induced by quantization, resulting in suboptimal allocation and routing. In this work, we propose Global Expert-level Mixed-precision Quantization (GEMQ) to overcome these limitations via (1) a global linear-programming formulation that captures model-wide expert importance based on quantization error analysis, and (2) efficient router fine-tuning to adapt routing to quantized experts. These components are integrated into a progressive quantization framework that iteratively refines importance estimation and allocation. Experiments demonstrate that GEMQ significantly reduces memory and accelerates inference with minimal accuracy degradation. Source code is available at https://github.com/jndeng/GEMQ .
Original Article
View Cached Full Text

Cached at: 05/25/26, 08:59 AM

# GEMQ: Global Expert-Level Mixed-Precision Quantization for MoE LLMs
Source: [https://arxiv.org/abs/2605.23078](https://arxiv.org/abs/2605.23078)
[View PDF](https://arxiv.org/pdf/2605.23078)

> Abstract:Mixture\-of\-Experts Large Language Models \(MoE\-LLMs\) achieve strong performance but incur substantial memory overhead due to massive expert parameters\. Mixed\-precision quantization mitigates this cost by allocating expert\-wise bit\-widths based on their importance, approaching the accuracy\-memory Pareto frontier and enabling extreme low\-bit quantization\. However, existing methods rely on layer\-wise importance estimation and overlook router shifts induced by quantization, resulting in suboptimal allocation and routing\. In this work, we propose Global Expert\-level Mixed\-precision Quantization \(GEMQ\) to overcome these limitations via \(1\) a global linear\-programming formulation that captures model\-wide expert importance based on quantization error analysis, and \(2\) efficient router fine\-tuning to adapt routing to quantized experts\. These components are integrated into a progressive quantization framework that iteratively refines importance estimation and allocation\. Experiments demonstrate that GEMQ significantly reduces memory and accelerates inference with minimal accuracy degradation\. Source code is available at[this https URL](https://github.com/jndeng/GEMQ)\.

## Submission history

From: Jianing Deng \[[view email](https://arxiv.org/show-email/49ac9fb1/2605.23078)\] **\[v1\]**Thu, 21 May 2026 22:23:38 UTC \(653 KB\)

Similar Articles

GEM: Geometric Entropy Mixing for Optimal LLM Data Curation

arXiv cs.LG

GEM reformulates LLM data curation as a variational problem on the hypersphere, using geometric entropy mixing and a minorize-maximize algorithm to discover balanced semantic clusters, achieving state-of-the-art improvements in data mixing strategies by up to 1.2% average downstream accuracy.

Mix-Quant: Quantized Prefilling, Precise Decoding for Agentic LLMs

arXiv cs.CL

Mix-Quant proposes a phase-aware quantization framework for agentic LLMs, using NVFP4 quantization for the prefilling stage to accelerate computation while preserving BF16 precision for decoding to maintain accuracy. The method achieves up to 3x speedup in prefilling with minimal performance degradation on agentic benchmarks.