Meituan open-sources LongCat-2.5, a 1.6T multimodal AI model

Meituan's LongCat team has open-sourced LongCat-2.5, a native multimodal model built for long-running AI agent tasks. It uses the MIT license, so anyone can deploy it for business or train it further without extra approval. The model has 1.6 trillion total parameters, with about 48 billion active per token, and supports a one-million-token context. It performs better on complex coding projects, visual coding, and office work, and can handle creative design and 3D tasks. The gains come from a new way of handling vision: images are turned into multi-level discrete tokens and processed by the main model just like text, with a visual codebook trained together with the main model. The team also tuned it for Chinese AI chips with limited memory and bandwidth, using a W4A4 method that compresses weights and activations to 4 bits to cut costs.