Meituan LongCat-2.0 synchronously opens up domestic card inference code According to a report by Golden Finance on July 9th, LongCat-2.0, as the industry's first trillion parameter model to complete inference on a 50000 card domestic computing power cluster, has been fully open sourced. We have conducted deep collaborative optimization in model architecture, chip adaptation, and deployment strategy for domestically produced computing chips with limited video memory and bandwidth, enabling the trillion parameter model to run steadily and quickly on existing cards. The open source of LongCat-2.0 is both a technical path disclosure and an ecological invitation. This open source synchronization provides high-precision versions such as BF16, FP8, and INT8, fully covering the deployment requirements of different computing power platforms. At the same time, we deeply embrace the open source community and will synchronously open source the inference results optimized for domestic computing power. This means that even without the latest computing power, LongCat-2.0 can still run stably based on existing hardware. We hope to use this out of the box inference stack to enable more domestic cards, including old cards, to smoothly deploy trillion dollar model inference services and unleash greater value in real productivity scenarios.