en.Wedoany.com Reported - On June 30, Meituan officially released its next-generation trillion-parameter large model, LongCat-2.0, and announced its open-source availability. The model has a total parameter count of 1.6T, with an average of approximately 48B activated parameters, a dynamic activation range of 33B to 56B, and native support for a 1M ultra-long context. LongCat-2.0 is a trillion-parameter model that underwent full-process training and inference on a domestic computing cluster of 50,000 GPUs, with pre-training data exceeding 30T tokens, covering Chinese, English, multilingual, and code data. It provides foundational model capabilities for long-text comprehension, code processing, multilingual tasks, and complex agent applications.
The release of LongCat-2.0 extends Meituan's large model capabilities from business applications to open-source foundational models. The 50,000-GPU domestic computing cluster, 1.6T total parameters, and 1M context window are the most notable technical highlights of this release.
In terms of model architecture, LongCat-2.0 adopts a configuration of trillion-level total parameters and tens of billions of activated parameters, with an average of approximately 48B activated and a dynamic activation range of 33B to 56B. Complex tasks can invoke more parameters, while lightweight tasks can reduce computational consumption. This design helps control inference costs and improves resource utilization efficiency across different task scenarios. The 1M ultra-long context capability allows the model to process larger documents, codebases, contract materials, project files, and multi-turn task records in a single pass, reducing information loss caused by segmenting long content.
Domestic computing power is another key aspect of this release. LongCat-2.0 completed training and inference on a 50,000-GPU domestic computing cluster, indicating that domestic AI infrastructure has entered the pipeline for ultra-large-scale model training.
Following its open-source release, LongCat-2.0 will be available to developers, enterprises, and research institutions. Industries such as manufacturing, retail, logistics, supply chain, and engineering services have extensive long-text materials, including equipment manuals, technical specifications, procurement contracts, project documents, customer service records, code repositories, and enterprise knowledge bases. The 1M context model can be used for knowledge base Q&A, long-document analysis, code assistance, business process automation, and agent task orchestration, lowering the barrier for enterprises to build industry-specific models and private applications.
Meituan's release of LongCat-2.0 signals its continued investment in large model foundational capabilities. The subsequent application effectiveness will still depend on model weights, technical reports, license terms, inference costs, and the scope of open-sourced supporting tools.
