LongCat-Flash-Chat
Meituan · 2025-08-29 · 561.9B parameters
LongCat-Flash-Chat is the open-weight, non-thinking text checkpoint of the LongCat-Flash family, built as a 28-layer sparse MoE with zero-computation experts and shortcut-connected MoE blocks. The released checkpoint contains 561,862,858,752 learned parameters and dynamically activates 18.6B to 31.3B parameters per token (about 27B on average); it supports a 128K-token context and is tuned for instruction following, coding, and agentic tool use.