Meituan open-sources LongCat-2.0, a 1.6-trillion-parameter MoE model with a 1-million-token context window
The model is optimized to run on Ascend NPU clusters.
Story Brief
The model is optimized to run on Ascend NPU clusters.
The model is optimized to run on Ascend NPU clusters.
The model is optimized to run on Ascend NPU clusters.
Fascinating. Wait, did LongCat train a 1.6T MoE on 910*B* NPUs, not even on CloudMatrix 384 with 910C's? That's a wholly ngmi hardware. Wow. "Tested on the Ascend A2" is not bulletproof evidence, they could just be considerate of end users… but I think yeah, mere Atlas A2. 2023. https://x.com/teortaxesTex/status/2074220864103448883/photo/1 https://twitter.com/JordanNanos/status/2073778641603805619