The market for models with open weights is growing: Meituan has released the weights of LongCat-2.5, a model with 1.6 trillion parameters designed for use in AI agents.Open weights means that the trained model parameters can be downloaded and run on your own infrastructure – in contrast to models that are only accessible via a provider's API.
Two things are noteworthy: the measured performance on coding tasks and the fact that training took place entirely on Chinese hardware.
The number that matters
On Terminal Bench, a benchmark for command-line tasks, LongCat-2.5 scores 88.4 points according to Meituan.This puts the model just behind Claude Opus 5.For a model whose weights are freely available, that is a narrow gap to the commercial top.
Terminal Bench tests whether a model actually completes multi-step tasks in a shell environment: creating files, executing commands, interpreting errors, refining them.This comes closer to what a coding agent must deliver in everyday project work than classic question-and-answer tests.
One limitation remains, as with all benchmark figures: they come from the provider and describe standardised tasks.Whether a model performs similarly well in a concrete codebase with grown structures, internal conventions and incomplete documentation cannot be derived from this.For this question, only your own test with your own tasks helps.
The catch lies in operation
Open weights do not automatically mean that a model is easy to operate.With LongCat-2.5 this becomes clear: the example configuration documented by Meituan runs on 194 Huawei Ascend chips.Notes on operation on Nvidia GPUs are missing.
This has practical consequences.Anyone who has so far planned with Nvidia hardware – i.e. with what is available in European data centres and at most cloud providers – will not find a ready-made guide.Whether and with what effort the model can be ported is not apparent from the release.
For small and medium-sized enterprises, operating a model of this magnitude themselves is not a realistic option anyway.Almost two hundred accelerator cards are not a server rack but a data centre investment.The relevant path therefore tends to lead via hosting providers that offer such models as a service – provided that providers can be found that operate the necessary hardware.
Why open weights still count
Even if self-operation is out of the question, a release like this changes the starting position.Three points are interesting from a project perspective:
- Negotiating position. The more high-performance models with open weights are available, the less a project is tied to a single provider's pricing and product policy.
- Choice of location. Models with open weights can in principle be operated wherever it suits legally and organisationally.This is an argument for data that should not leave the company – provided a suitable operator is available.
- Planning certainty. A model whose weights are available cannot be switched off overnight or changed silently.Both have happened with API models in the past.
At the same time, the limits should be clearly named.Open weights are not the same as open source: training data and training code are not necessarily disclosed, and the licence terms can restrict commercial use.Anyone who wants to use such a model productively should have the licence checked in advance.
What the hardware question signals
The fact that LongCat-2.5 was trained entirely on Chinese hardware is a signal both technically and strategically.It shows that model training at this scale no longer relies exclusively on one manufacturer's ecosystem.
For project managers, this above all means a monitoring task.If model ecosystems differentiate along different hardware platforms, the question "Does this run on the infrastructure we can use?" becomes a selection criterion in its own right – alongside performance, cost and licence.
A pragmatic approach
For companies planning AI-supported development, such releases do not create pressure to act, but they do suggest an architectural recommendation:
- Abstract model access. If applications are not programmed directly against a single provider's interface but against your own intermediate layer, a switch remains a configuration question instead of a rebuild.
- Build your own test cases. A small collection of typical tasks from your own everyday project work says more about a model's suitability than any benchmark table.
- Separate requirements. Not every task needs the strongest available model.Anyone who cleanly separates task types and model classes lowers costs and reduces dependence on a single top model.
Classification for project practice
For web and software projects, LongCat-2.5 shifts less the tool selection of today than the expectation of tomorrow: the performance gap between open and closed models on coding tasks is getting smaller, and that strengthens the position of those who formulate data sovereignty and provider independence as a requirement.In the short term, however, operating such models remains a matter for specialised providers due to the hardware requirements.
The sensible consequence is therefore not an immediate switch, but an architecture that allows a switch later without major effort.