Skip to content
BAAI Community·Indexed by this site · Originally published· AI updates →· Its place in the scan record →

Meituan open-sources the 1.6 trillion parameter MoE agent model LongCat-2.5, with self-hosting offering only an Ascend solution

美团开源 LongCat-2.5 权重:1.6 万亿参数 MoE Agent 模型,自托管指南未提供 Nvidia GPU 路径 - 智源社区

AI summary

Meituan's LongCat team has open-sourced the weights of LongCat-2.5 under the MIT license, permitting commercial use, modification and redistribution; it is a MoE model with 1.6 trillion total parameters, activating about 48 billion parameters per token, positioned as an agent model for long-horizon multi-step tasks such as coding. The official self-hosting guide is based on a 194-card Huawei Ascend Atlas A2 cluster and the CANN software stack, using separate deployment of Encoder, Prefill and Decode, and gives no deployment guidance for Nvidia GPUs. The model applies 4-bit quantization to expert-layer weights and activations (W4A4), and Meituan says that compared with the W8A8 version, prompt processing throughput improves by 43.3% and generation throughput by 42.77%.

Read originalhub.baai.ac.cn