LongCat-2.0: 1.6 T‑parameter MoE model trained on Chinese ASICs
Meituan released LongCat‑2.0, a 1.6‑trillion‑parameter mixture‑of‑experts LLM that activates only ~48 B parameters per token. The entire training run used Meituan’s proprietary AI ASIC superpods, marking a rare shift from GPU‑dominated large‑scale model training and showing China’s push for hardware independence.