On July 9th, LongCat-2.0, the industry’s first trillion-parameter model to successfully complete inference on a cluster of 50,000 domestically produced computing cards, was officially open-sourced. The model has undergone extensive optimization in terms of model architecture, chip compatibility, and deployment strategies for domestic computing cards, which often have limited video memory and bandwidth. These optimizations ensure the stable and efficient operation of the trillion-parameter model on existing hardware infrastructure.

The open-sourcing of LongCat-2.0 not only reveals the technical approach behind its development but also serves as an invitation to ecosystem partners to collaborate and innovate. This release includes multi-precision versions of the model, such as BF16, FP8, and INT8, to cater to the diverse deployment requirements of various computing platforms.

Furthermore, the project team is actively engaging with the open-source community by simultaneously releasing inference results that have been optimized for domestic computing power. This ensures that users can reliably run LongCat-2.0 on their existing hardware, even without access to the latest computing resources. The team aims to facilitate the seamless deployment of trillion-parameter large model inference services across a wider range of domestic computing cards, including older models, through this ready-to-use inference stack. By doing so, they hope to unlock greater value for these cards in real-world productivity scenarios.