Meituan LongCat Flash

LongCat Flash Thinking is a 560B-parameter MoE reasoning model with 512 experts, activating 18.6-31.3B parameters per token. It uses a 28-layer transformer with 6,144 hidden size, 64 attention heads, and 131K context length. The design includes zero-computation experts and MLA attention. Trained via a two-phase pipeline, Long CoT cold-start and large-scale RL on the DORA system, it emphasizes formal reasoning, theorem proving, and agentic tool use, with domain-parallel RL improving stability and cross-domain performance.

LongCat Flash Chat is a 562B-parameter Mixture-of-Experts (MoE) language model optimized for agentic tasks and high-throughput inference. It dynamically activates 18.6-31.3B parameters per token using zero-computation experts and a Shortcut-connected MoE (ScMoE) design. Built on a 28-layer transformer with 64 attention heads and 6,144 hidden size, it supports up to 128K token context. The model employs a multi-stage training pipeline with reasoning-focused pretraining, agentic post-training, and multi-agent task synthesis, enabling advanced reasoning, coding, and iterative interaction capabilities.