Meituan releases LongCat-2.5, 1.6T multimodal model for agentic coding

A 1.6T MoE built for long-horizon coding agents shows Chinese labs pushing agentic models.

LongCat-2.5 is a 1.6T-parameter native multimodal model with about 48B active per token, trained with long-horizon reinforcement learning for agentic coding and optimized for W4A4 inference on domestic Chinese compute.

Name

LongCat-2.5

Params

1.6T total, ~48B active

Domain

software

Read it at Pandaily →

Coverage: 2 sources, in order

1. 02 Nov 02:10 Pandaily Read the article →
2. 02 Nov 04:32 TechNode Read the article →

← Front page