LONGCAT-2.0 LAUNCHES WITH 1.6T PARAMETERS

The model uses a mixture-of-experts approach, activating only 48B of its 1.6T parameters during inference. This design reduces computational overhead compared to dense models of equivalent scale while maintaining performance. MoE architectures route inputs to specialized sub-networks, allowing larger effective model capacity without proportional increases in active computation. LongCat-2.0 follows this pattern to enable faster inference and lower resource requirements. The release generated 147 points and 43 comments on Hacker News, indicating developer interest in efficient large-scale models. More details are available on LongCat's blog and discussion thread. The model represents ongoing efforts to scale language models efficiently as organizations balance performance requirements against computational costs.