Meituan's LongCat team open-sourced the multimodal large model LongCat-Next

Meituan's LongCat team has launched the LongCat-Next multimodal model, with its core innovation being the LoZA sparse attention mechanism. The model supports ultra-long contexts of up to 1MB and intelligently replaces 50% of low-importance modules with streaming sparse attention, achieving a 10x speedup in 128K decoding and a 30% reduction in computational power. The model includes two versions, Flash-Exp and Flash-Lite, and outperforms Qwen-3 in long text tasks.