Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

What happened

Meituan has released LongCat-2.5, a 1.6-trillion-parameter AI agent model, making its weights publicly available one week after the model's official unveiling. The model was trained entirely on Chinese hardware, specifically Huawei Ascend chips, with no use of Nvidia GPUs. LongCat-2.5 is designed for autonomous multi-step tasks, especially coding, and supports multimodal inputs including text, images, and video. The model is accessible via API and a chat interface, and is available under an MIT license.

Key details

• LongCat-2.5 uses a mixture-of-experts architecture, with about 48 billion parameters active per token.
• The model can process up to one million tokens at once and generate responses up to 128,000 tokens long.
• A sample deployment uses 194 Huawei Ascend accelerators, with no support for Nvidia GPUs.
• On the Terminal Bench v2.1 coding benchmark, LongCat-2.5 scores 88.4, close to Claude Opus 5's 89.1.
• The weights are released under an MIT license, allowing commercial use and modification, and the API charges about €0.27 per million input tokens and €1.07 per million output tokens during the promotional period.

Why it matters

Chinese AI developers and companies now have access to a large-scale, open-weight agent model that runs efficiently on domestic hardware, reducing reliance on Nvidia GPUs. This enables organizations in China and elsewhere to deploy advanced AI capabilities using Chinese chips, potentially lowering costs and sidestepping export restrictions. Developers benefit from the MIT license, which permits broad commercial and research use, and can integrate LongCat-2.5 into existing agent frameworks and applications.