Meituan releases LongCat-2.5, a 1.6-trillion-parameter multimodal model for long agentic coding tasks
The model activates about 48 billion parameters per token and was trained on real agent failures turned into reproducible tasks.
Chinese delivery company Meituan released LongCat-2.5 on November 2, adding stronger visual understanding and coding to its model lineup. The model has 1.6 trillion total parameters, with about 48 billion activated for each token, and is built for tasks that AI agents must carry out over long periods. Meituan said it expanded training on long task sequences and used large-scale reinforcement learning to improve the model's work on complex codebases and complete software projects. It also converts failures that real agents ran into into reproducible training tasks, to close gaps in the model's performance. LongCat-2.5 handles images and text in a single token framework, which the company says supports visual coding, creative design and 3D tasks. The model is available through Meituan's online chat service and API platform, and releases on GitHub and Hugging Face are planned. The report includes no benchmark figures, so it is not yet possible to say how LongCat-2.5 compares with other open or proprietary coding models. The release adds to a series of large Chinese models built specifically for agentic software engineering.