Meituan Releases LongCat-2.5 — A 1.6-Trillion-Parameter Open-Weight Model Under the MIT License

Another trillion-parameter model just went open weight — and this one's licence is about as permissive as they come. Meituan has released LongCat-2.5, a mixture-of-experts model with roughly 1.6 trillion parameters in total and around 48 billion active on each token, under the MIT license. Anyone can use it, change it and build on it, including for commercial work.

What LongCat-2.5 is built for

LongCat-2.5 is aimed squarely at agents. Meituan describes it as a native multimodal mixture-of-experts model built for long-horizon agent tasks — the kind of work where a model has to keep going for a long stretch, pulling in different kinds of input along the way. It builds on the earlier LongCat-2.0, and the company says it marks a first step toward systems that can handle "superhuman-scale" agent work. The sparse design is the trick: most of the model's 1.6 trillion parameters stay dormant on any given token, with only about 48 billion switching on.

What the MIT licence changes

Plenty of open-weight models come with strings attached — custom licences that limit commercial use or demand a revenue share. An MIT licence carries none of that. Builders can fine-tune LongCat-2.5, ship it in a product and charge for it without asking Meituan first. Meituan's announcement points to the model on Hugging Face, where the weights, config and licence are all listed.

The parameter count, in context

Is 1.6 trillion parameters a lot? Yes — but the number that matters for anyone running the model is the 48 billion active per token. That's the slice that actually does the work, so running costs land closer to a much smaller model. Sparse models like this one are common now, and the race is over who can squeeze the most ability out of each active parameter. Google's new TPU 8tx chips are built for exactly this kind of agentic AI — and the same week saw DeepMind's AI formally verify Signal's crypto protocol. The models are getting bigger, the proofs are getting harder, and the licences are getting looser.

Whether LongCat-2.5 finds its way into real products is the test that counts — and with an MIT licence, there's nothing stopping anyone from trying.

Originally published on Rohan Tech Official.