Meituan Open-Sources 1.6T-Parameter LongCat-2.5 Multimodal Model

Meituan's LongCat team has officially open-sourced LongCat-2.5, a native multimodal large model with 1.6 trillion total parameters and approximately 48 billion activated per token. The model supports a million-token context window and is released under the MIT license, permitting commercial inference and further training. This marks one of the largest open-source models in China to date and signals Meituan's deeper push from application layers into foundational AI infrastructure.

Technically, LongCat-2.5 quantizes visual signals into multi-level discrete tokens that share backbone parameters with text, rather than relying on a separate vision module, achieving true native multimodal fusion. It also employs W4A4 4-bit quantization to better fit domestic AI chips with limited memory and bandwidth, lowering deployment barriers. Third-party coding platform OpenCode has simultaneously launched the model with a two-week free access window, highlighting its 1M-token context, multimodal input, and zero data retention, though official benchmark scores and pricing remain undisclosed.

The open-source release carries significant ecosystem implications: the MIT license grants extensive freedom for commercial use and derivative training, likely attracting developers to build vertical applications on LongCat-2.5. Its native multimodal architecture and domestic-chip adaptation offer a reusable technical path for China's compute-constrained AI industry. By entering the open-source race at the 1.6-trillion-parameter scale, Meituan could reshape the competitive landscape of leading open models and accelerate the transition of multimodal capabilities from lab to large-scale deployment.