Meituan Open-Sources LongCat-Flash-Lite: 69B Total Params with 3B Active

teortaxesTex · x · 2026-08-01

Meituan's tech team has released the new LongCat-Flash-Lite-Sparse model on Hugging Face. The model utilizes an ultra-sparse MoE (Mixture of Experts) architecture with optional Hierarchical Indexing.

It has a total parameter count of 69B but only activates approximately 3B parameters during inference. This design aims to significantly reduce actual inference compute costs, making it a highly interesting architectural experiment.

Related event: Meituan Open-Sources LongCat-Flash-Lite-Sparse Model(2 posts)→

Original post →

More from Models

Models channel →