Meituan open-sourced LongCat-2.5 on November 10 under the MIT License. It supports commercial inference deployment and further training without additional authorization. The model targets long-horizon agent tasks and features a native multimodal architecture, with 1.6 trillion total parameters, about 48 billion parameters activated per token, and support for a one-million-token context. It significantly improves end-to-end delivery capabilities in complex codebase development, complete project construction, Visual Coding, and professional office scenarios. The weights are available on HuggingFace, with the technical blog and API platform also opened simultaneously.
The capability improvements stem from scaling the entire training pipeline. The visual representation layer represents visual signals as multi-level discrete tokens, which are processed uniformly by the backbone together with text rather than by an external vision module. It performs residual vector quantization on native-resolution features, progressively filling in details level by level. The visual codebook is trained jointly with the backbone, and visual tokens can be cached offline. On the professional office side, it strengthens material processing and professional judgment, enabling end-to-end delivery of reports, presentations, and office tools.
Inference adopts the W4A4 scheme, quantizing expert weights and activations to 4-bit and executing them on native INT4 matrix multiplication units. FlatQuant is used to optimize MoE quantization accuracy, with downstream performance comparable to that before quantization, while dedicated fused operators help reduce costs. Both training and supervised fine-tuning use long-sequence training. Targeting domestic AI accelerators with less than 80GB of VRAM per card, it uses FSDP double sharding and hybrid vocabulary parallelism to distribute memory pressure. This configuration increases per-GPU throughput in long-sequence SFT by 86% over the DDP baseline, providing an efficiency foundation for open-source deployment of trillion-parameter MoE models.
Meituan stated that the MIT License is intended to support free use, modification, and redistribution of the model, and that it will continue to open-source models and engineering experience. Running a trillion-parameter MoE under memory and bandwidth constraints continues the trend in large-model competition shifting from parameter scale toward engineering implementation and adaptation to domestic compute.
