ConMoE framework compresses MoE models without retraining

Researchers have developed ConMoE, a novel framework for compressing Mixture-of-Experts (MoE) language models without requiring retraining. This method consolidates the expert pool by reassigning original expert references to a smaller set of selected prototypes. ConMoE uses calibration-based signals to choose which experts to retain and how to remap calls, preserving the original router interface. Experiments on models like deepseek-moe-16b-base and Qwen3-30B-A3B demonstrate that ConMoE achieves competitive or superior compression rates compared to existing pruning and merging techniques. AI

IMPACT This research offers a method to reduce the memory footprint of MoE models, potentially making them more accessible for deployment.

RANK_REASON This is a research paper detailing a new method for compressing MoE models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ConMoE
• deepseek-moe-16b-base
• Mixture-of-Experts
• OLMoE-1B-7B-0125
• Qwen3-30B-A3B

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →