Chinese Meituan's LongCat-2.5: A massive 1.6-trillion-parameter AI agent model built for coding, trained entirely on China's own chips.

Chinese food delivery and tech giant Meituan has open-sourced the weights for its LongCat-2.5 model, a massive 1.6-trillion-parameter AI system designed for autonomous agents that handle complex, multi-step tasks like coding. Built as a mixture-of-experts model, it activates only about 48 billion of its total parameters for each token processed. It handles up to a million tokens of input and can generate responses up to 128,000 tokens long, with multimodal capabilities allowing it to understand images, video, and text.

This represents a significant technical achievement in AI development, particularly notable for being trained entirely on Chinese hardware like Huawei chips instead of Nvidia GPUs. The model's performance reportedly approaches top Western models like Claude Opus 5 on coding benchmarks, showcasing China's growing independence in the AI hardware and software stack. The model's design for agentic work—using real GitHub tasks and blocking training shortcuts—points toward practical applications in automated software development and complex problem-solving.