New LorExperts and BTExperts methods compress MoE models effectively

Researchers have developed two new methods, LorExperts and BTExperts, for compressing Mixture-of-Experts (MoE) language models. These techniques aim to reduce the computational cost of deploying MoE models by compressing expert weight matrices while preserving accuracy. LorExperts clusters experts and uses a dominant full-precision expert with low-rank corrections for others, retaining the original router. BTExperts further organizes these into a tree structure for amortized computation. Both methods show improved performance over existing baselines, particularly as the number of experts increases, demonstrating their effectiveness in maintaining accuracy with significant compression on models like Qwen3-30B-A3B and Gemma-4-26B-A4B. AI

IMPACT These compression techniques could significantly reduce the cost of deploying large MoE models, making advanced AI more accessible.

RANK_REASON The cluster contains a research paper detailing new methods for compressing MoE language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• BTExperts
• D^2-MoE
• Gemma 4 26B-A4B
• Inesh Chakrabarti
• LorExperts
• mixture of experts
• Qwen3-30B-A3B
• Reap

• paper
• infra

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