Meituan's LongCat team released the model weights for LongCat-2.5 under the MIT license, marking a significant milestone as the 1.6 trillion-parameter mixture-of-experts (MoE) model was trained entirely on Chinese domestic hardware. The model activates approximately 48 billion parameters per token and supports a 1-million-token context window.
Benchmark results show LongCat-2.5 scoring 88.4 on Terminal Bench v2.1, placing it just behind Claude Opus 5 (89.1) and GPT-5.6 Sol (88.8). Meituan also published a 4-bit quantized version with a self-hosting guide built on 194 Huawei Ascend chips — notably offering no NVIDIA GPU option. Limited-time API pricing stands at roughly €0.27 per million input tokens and €1.07 per million output tokens.
Meituan Open-Sources 1.6T-Parameter LongCat-2.5 Model Trained Entirely on Domestic Chinese Hardware
Disclaimer: The content provided on Phemex News is for informational purposes only. We do not guarantee the quality, accuracy, or completeness of the information sourced from third-party articles. The content on this page does not constitute financial or investment advice. We strongly encourage you to conduct you own research and consult with a qualified financial advisor before making any investment decisions.