LongCat-Flash-Lite
Meituan · 2026-02-05 · 69.1B parameters
LongCat-Flash-Lite is Meituan's non-thinking sparse mixture-of-experts language model, published as a 68.5B-parameter model with 2.9B to 4.5B parameters dynamically activated per token. Trained from scratch, it allocates 31.4B parameters to an N-gram embedding module rather than additional experts to improve model performance and inference efficiency. It supports contexts up to 256K tokens using YaRN and targets agentic tool use and coding.