Meituan released its next-generation large model LongCat-2.5 on November 2, according to multiple reports from Eastmoney and Securities Star. The natively multimodal model, built for long-horizon agent tasks, carries 1.6 trillion parameters and was trained end-to-end on domestic compute.
The groundwork was laid in June, when LongCat-2.0 completed full-cycle training and inference of a trillion-parameter model on domestic compute for the first time. Version 2.5 strengthens native visual understanding and agent coding on that base, converting visual input into tokens the model processes uniformly to support tasks such as interactive web pages and 3D showcases. The company's release notes say overall capability now matches mature international closed-source models and leading open-source models.
Engineering adaptation on the training side is the key increment this time. Mid-training used sequences up to 1M tokens, with cumulative training volume in the trillions of tokens; when multiple cards process long sequences together, memory capacity and interconnect bandwidth directly determine efficiency. Meituan worked with domestic compute vendors to optimize task partitioning around supernodes, keeping parameters and task data on-card as much as possible. Its technical blog reports single-card training throughput up 86% versus a DDP (DistributedDataParallel) baseline — one card doing the work of about 1.86 cards.
Inference makes a similar quantitative trade-off: the W4A4 scheme compresses both expert weights and activations to 4 bits and tunes computation and transport for the hardware's traits, holding performance roughly equal to the pre-quantization model on evaluated downstream tasks.
On purpose, Meituan's answer remains supporting its core business. CEO Wang Xing said at the Q2 earnings call, "We will not become a token factory," pledging that models and AI products will serve users, merchants and operating efficiency. The AI operations workbench "Miaoshou," built on the LongCat series, already serves chain brands including Xu Sheng Ji and Yihetang across brand analytics, store management and customer acquisition.
From 2.0 to 2.5, Meituan has placed its model validation inside its own business workflows. Whether the "domestic chips, domestic models" route keeps delivering will be seen in merchant-side adoption and effect data.