The model utilizes a mixture-of-experts architecture with 1.6 trillion total parameters and 48 billion active parameters. Meituan omitted performance benchmarks for this launch, whereas the previous LongCat 2.0 version was trained on Ascend 910B chips and claimed state-of-the-art results. Some developers who tested LongCat 2.0 in small-scale evaluations described the predecessor as completely hopeless.
