Meituan launches LongCat-2.5 AI model with 1.6 trillion parameters
The multimodal model activates about 48 billion parameters per token and targets long-running AI agent work.
The 30-second read
- Meituan released the multimodal LongCat-2.5 model on November 2.
- It has 1.6 trillion total parameters, with about 48 billion active per token.
- It is live via online chat and API; GitHub and Hugging Face releases are to follow.
IllustrationMeituan released LongCat-2.5, a multimodal AI model, on November 2, according to TechNode. TechNode first reported the release date. The model has 1.6 trillion total parameters. About 48 billion of them are active for each token.
For training, the model was trained longer on long task sequences. It also went through large-scale reinforcement learning aimed at complex codebases and full software projects. First reported by TechNode: failures seen in real agents are turned into training tasks that can be run again.
TechNode also first reported that the model handles images and text in a unified token framework. It supports visual coding, creative design and 3D tasks. It is available through online chat and an API platform, and GitHub and Hugging Face releases are to follow. Pandaily first reported that it is optimised for W4A4 inference on domestic compute.
The model is built for tasks where AI agents must keep working over long stretches. It adds to Meituan's existing lineup of AI models.
Why it matters
Indian developers and startups building AI agents can watch whether the promised GitHub and Hugging Face releases arrive. Open access would let them test the model themselves.
