The promise from November 2 is now a download
When LongCat-2.5 launched on November 2, the release stopped at Meituan's own doors. Anyone could use the chat service or the API. The GitHub and Hugging Face releases were described as planned, and until they existed, the launch claims could not be checked against the weights themselves.
That part has changed. The weights are on Hugging Face, ungated, under an MIT license. MIT is the permissive end of the scale: commercial use, modification and redistribution are allowed under the license terms, with no revenue ceiling and no territory carve-out of the kind that has followed some other open video and language releases this year. Whatever else is said about this model, the license itself is not the catch.
Count the files, not the headline
The model card repeats the launch headline: 1.6 trillion total parameters, about 48 billion active per token. The repository's own metadata tells a slightly different story. ALFW's sum of the published safetensors metadata gives roughly 1.78 trillion parameters across 210 files, in a repository of about 3.56 terabytes as shipped.
The two figures may not be counting the same things. Vision components and embedding tables are exactly the parts a headline count sometimes leaves out, and the card does not say how its 1.6 trillion was counted. Nothing Meituan has published reconciles the two numbers. But the practical consequence does not depend on the explanation: anyone sizing hardware, storage or a serving budget from the card's number is working from a figure about 10 percent below the files they actually have to hold (ALFW arithmetic: the card's figure sits roughly 10.1 percent under the metadata total). At this scale, that is not rounding.
This was the open question in the November 2 coverage, in a different form. Then, the 1.6 trillion figure looked carried over from LongCat-2.0's model card. Now Meituan has published a card of its own for 2.5, and the number still does not match its own files. The weights being public is what makes that checkable at all, which is the strongest argument the release makes for itself.
Open weights, one hardware map
The second document worth reading is the serving guide. The model card's deployment section says the model can be deployed on NPU platforms, and links Meituan's SGLang-FluentLLM guide. That guide documents a sample self-hosting configuration built on 194 Huawei Ascend Atlas A2 accelerators, split into 2 chips for the encoder, 64 for prefill and 128 for decode, running real weights on Huawei's CANN platform. The quantized build uses 4-bit weights and activations, a precision Meituan notes the domestic chips handle natively. There is no corresponding Nvidia deployment guide in the card's deployment section, and nothing the desk reviewed establishes a minimum deployment size; 194 cards is the demonstrated configuration, not a stated requirement.
Read that as a statement about who this release is for. A team already running Ascend infrastructure gets a documented, vendor-tested path. A team on Nvidia hardware gets the weights and a porting project. The MIT license removes the legal barrier to doing that port. It does not remove the engineering cost.
None of this makes the release meaningless. A 1.6-trillion-class agent model under MIT is a real data point about how far open releases have come. But the honest description is narrower than the headline version: the weights are open, the license is permissive, and the documented way to run them sits inside a single domestic hardware ecosystem.
What has not changed
The scores are still Meituan's. Terminal Bench v2.1 at 88.4 and the rest of the launch table were vendor-run in November, and the sources reviewed for this piece provide no independent reproduction of them. The API remains the realistic route for almost everyone: secondary reports transcribing LongCat's own platform pricing list put the limited-time rate at $0.30 per million input tokens, and the API runs on Meituan's hardware rather than a cluster of your own.
So the order of operations is the same one this site suggested at launch, with one step now actually available. Use the API to test the claims. If the model earns a place in your stack, read the repository files rather than the card before believing the parameter count, and read the serving guide before believing that open weights mean you can run it.
Sources
- [1] Hugging Face — meituan-longcat/LongCat-2.5 repository and model card (license, repository dates, file metadata, deployment section)Read source
- [2] Meituan — SGLang-FluentLLM NPU deployment guide (Atlas A2 194-card sample configuration)Read source
- [3] Unite.AI — LongCat-2.5 launch report transcribing the LongCat API platform's pricing listRead source
- [4] TeamoRouter — LongCat-2.5 API price and specs, from the LongCat API PlatformRead source