By Hamza Faisal

Meituan has released the full weights of LongCat-2.5, its flagship artificial intelligence model, giving developers free rein to download, run and build on one of the largest open-weight systems ever published. The Chinese food-delivery and local-services giant published the checkpoint on Hugging Face on November 9, a week after introducing the model, under the MIT licence — the most permissive widely used open licence, allowing commercial use, modification and redistribution.

The numbers are striking. LongCat-2.5 is a mixture-of-experts model with about 1.6 trillion total parameters, of which roughly 48 billion are active for each token it processes. It is multimodal, reading screenshots, charts, images and video alongside text, and it supports a context window of one million tokens. Meituan describes it as a model built for long-horizon agent work: software that inspects code, revises it over multiple steps and handles office tasks without constant human prompting.

A giant model, a permissive licence

The MIT licence is the centre of the story. Unlike research-only or non-commercial releases, it lets companies deploy LongCat-2.5 in products, fine-tune it and redistribute it, provided they keep the licence notice attached. Meituan has also published a technical write-up and an API guide, and its hosted version is priced aggressively: a limited-time rate of 0.30 dollars per million uncached input tokens and 1.20 dollars per million output tokens, with cached input at 0.006 dollars.

That pricing is a deliberate challenge to hosted frontier models. Meituan is positioning LongCat-2.5 as a cheaper way to run agent workloads at scale, and the open weights mean teams are not locked into the company's own servers — at least in principle.

The catch: 3.5 terabytes and Chinese chips

In practice, running the model yourself is a serious undertaking. The checkpoint weighs about 3.56 terabytes in BF16 format, so self-hosting needs a cluster of servers with serious memory, not a workstation. Meituan's deployment guidance is also aimed squarely at domestic Chinese accelerators, with W4A4 quantisation instructions for neural processing units — a reminder that the release was built for China's own hardware stack.

There is even a small mystery in the numbers. Meituan's technical description says 1.6 trillion parameters, but the Hugging Face repository metadata lists 1.8 trillion — a gap probably explained by embedding tables the company does not count in its headline figure. Either way, it is a very large model by any measure.

Open weights, not open source

The release also comes with the usual caveats. The training data and the full training pipeline have not been published, so LongCat-2.5 is open-weight rather than fully open source. Its benchmark results — on software-engineering tests such as Terminal-Bench and DeepSWE — were measured by Meituan itself, and no independent verification has been published yet. VentureBeat also notes growing scrutiny in Washington over Chinese AI models, which could limit adoption by some Western teams.

Even so, the release is a significant moment for the open-weight movement. A Hong Kong-listed company best known for delivering food has now published one of the biggest openly usable AI models in the world — and under a licence that lets anyone use it commercially. The model debuted on November 2; the weights followed a week later. The test now is whether developers actually build on it.