Meituan makes 1.6T multimodal MoE model LongCat-2.5 generally available
LongCat-2.5 activates about 48B of its 1.6T parameters per token, processes visual signals alongside text as layered discrete tokens, and uses long-horizon reinforcement learning to improve work on complex codebases and full projects. Meituan reports that its W4A4 inference, which quantizes expert weights and activations to 4 bits, improved prefill and decode throughput by 43.3% and 42.77% over W8A8 in online load tests.
Visit original source