Meituan has published the weights of LongCat-2.5, a very large open-weights model, for anyone to download. The release is real and the license is permissive, but the files are far too big for ordinary hardware.

What Meituan released

Meituan, a Chinese company best known for food delivery, put LongCat-2.5 on Hugging Face, where the model card describes it as a native multimodal mixture-of-experts (MoE) model. In an MoE model, only part of the network runs for each token of text.

Per Meituan's model card, it has about 1.6 trillion total parameters and about 48 billion active per token. Meituan announced the release in a post on X.

Hugging Face's page metadata shows a different count: about 1.78 trillion parameters, split between BF16 and a small F32 portion. We did not reconcile the two figures.

The weights are under the MIT license. The card also says the MIT license "does not grant any rights to use Meituan trademarks or patents."

Who can actually run it

Almost no one outside a data center. Hugging Face's file listing shows the repository at 3.56 TB across 231 files, mostly BF16 weights. That is far beyond any laptop or single consumer GPU.

The card says the model can be deployed on NPU platforms and points to SGLang-FluentLLM. It gives no GPU or memory guidance, so we cannot say what hardware setup is needed.

What is not known

We did not find any statement in the model card on whether training code or training data are released. Treat both as not confirmed.

Meituan's benchmark scores are mostly measured in-house, and the card itself calls them "indicative only." We have not repeated any of them here. No independent verification of the model's performance has turned up.

The gap between the 1.6 trillion and 1.78 trillion parameter figures is also still open.

What this means for you

Most readers have nothing to download or change. This matters if you follow open models, since it is a very large set of weights under a permissive license. It is not a recommendation to run it.

What we checked

We read the Hugging Face model card and file listing on November 10. We link Meituan's X post as the announcement but did not verify it independently. We did not run the model or verify any benchmark results.

For related coverage of Chinese open-weight models, see the Qwen3.8 sandbox test from Cohere and DeepSeek's reported $1.5 billion funding round.