Meituan open-sources LongCat-2.5 multimodal model
Meituan released LongCat-2.5 under the MIT license, making the 1.6-trillion-parameter mixture-of-experts model available for commercial inference and further training. The native multimodal model activates about 48 billion parameters per token, supports a one-million-token context window, and is aimed at long-horizon agent tasks, including code development, Visual Coding, and professional office work. Meituan also published the weights, technical blog, API platform, W4A4 inference design, and adaptations for domestic AI accelerators; performance and throughput claims remain vendor-reported.
The release gives developers access to a million-context, natively multimodal model with open weights, permissive licensing, and documented deployment methods for both general and domestic-computing environments.