Meituan releases LongCat-Flash-Lite-Sparse MoE model

The model uses ~3B active parameters plus a 30B n-gram lookup table offloaded to RAM, targeting fast 256k context on a 24GB GPU. Early community analysis suggests it won't replace Qwen 3.6 27B.