Meituan LongCat launches LongCat-2.0, a 1.6T MoE flagship model built for agentic coding

美团 LongCat 发布旗舰模型 LongCat-2.0

Meituan LongCat released LongCat-2.0, a 1.6T-parameter MoE model with ~48B active parameters and native 1M context window. Built for agentic coding, it uses three techniques: LSA sparse attention for efficient 1M scaling, Zero-Compute Experts that dynamically activate 33B–56B parameters per token with no wasted compute, and MOPD which partitions experts into Agent, Reasoning, and Interaction groups with task-gated routing. It scored 59.5 on SWE-bench Pro, close to leading closed-source models. Pricing: input cache $0.015/1M tokens, input $0.75/1M tokens, output $2.95/1M tokens. Available now on SiliconFlow Day 0.

Why it matters: Meituan's LongCat drops a 1.6T-param flagship with ~48B active params and native 1M-token context, explicitly targeting agent workflows. The technical details (LSA sparse attention, Zero-Compute Experts) are concrete, not vaporware. Held back from 85+ because we only have the ...