A model designed to work across images and a whole coding project is now available for others to build on.
Meituan released LongCat-2.5 on November 9, publishing the model’s underlying files on Hugging Face under the permissive MIT license. That means developers can use and adapt the model itself, rather than only interact with a finished product.
The model is designed to connect visual understanding with multi-step work, such as developing a codebase, building a full project, or creating designs. The big idea is to help AI work on a task that unfolds over time, not just answer one question and stop.
Inside, it works like a large team of specialists. For each piece of text, it calls in only a small share of its parts, about 3%, while the rest stay idle.
That doesn’t make it easy for everyone to run. LongCat-2.5 is still a very large model, so teams need serious computing resources to use it themselves. But those with the infrastructure can inspect and adapt the model to their own work.
The weights are public; the real test is what independent teams can build with them. Can a model designed for whole projects become a dependable tool for getting those projects done?
