LongCat-2.0 launches with 1.6 trillion parameters and 48 billion active experts

LongCat-2.0 is introduced as a new large‑scale MoE model. It totals 1.6 trillion parameters across all experts. Only 48 billion parameters are active at inference time. The model builds on the original LongCat architecture. Its size aims to push the limits of language modeling. The MoE design balances capacity with computational cost. Developers can explore the model via the LongCat website. Future work may compare its performance to prior versions.