PANews|Nov 15, 2026 11:31
Meituan Open-Sources 1.6 Trillion-Parameter Intelligent Agent Model LongCat-2.5, Trained Entirely on Domestic Chips
According to The Decoder, Meituan's LongCat team has released the weights for the LongCat-2.5 model, licensed under the MIT license. This model uses a hybrid expert (MoE) architecture with a total of 1.6 trillion parameters, approximately 48 billion activations per token, and supports 1 million token contexts. It was trained entirely on Chinese hardware. Data released by Meituan shows that the model scores 88.4 on Terminal Bench v2.1, 89.1 on Claude Opus 5, and 88.8 on GPT-5.6 Sol. Meituan also released a 4-bit quantized version. The self-hosted version is based on 194 Huawei Ascend chips and does not offer an NVIDIA GPU solution. The API is priced at approximately €0.27 per million input tokens and €1.07 per million output tokens for a limited time.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink