Lates News

As the first in the industry to complete inference on a domestic computing power cluster of 50,000 cards, the trillion parameter model LongCat-2.0 has been fully open-sourced. In order to address the limitations of memory and bandwidth on domestic computing power chips, we have conducted deep collaborative optimizations on model architecture, chip adaptation, and deployment strategies, allowing the trillion parameter model to run stably and quickly on existing cards. The open sourcing of LongCat-2.0 is both a public display of technological progress and an invitation to the ecosystem. This open sourcing includes BF16, FP8, and INT8 versions that cover the deployment needs of different computing platforms with different precisions. At the same time, we deeply embrace the open-source community and will simultaneously open-source inference results that are optimized for domestic computing power. This means that even without the latest computing power, LongCat-2.0 can be stably deployed based on existing hardware. We hope that through this ready-to-use inference stack, more domestic cards, including older ones, can smoothly deploy trillion-scale model inference services and realize greater value in real productivity scenarios.