Meituan opens LongCat-2.5 weights under MIT
Meituan uploaded LongCat-2.5 weights to Hugging Face under MIT on November 9; the model card says 1.6T parameters, while file metadata reports about 1.78T.
For teams that want to inspect, fine-tune or host a model themselves, the news is a change in access: Meituan uploaded LongCat-2.5 weights to Hugging Face on November 9, 2026, with the repository marked MIT. LongCat-2.5 is Meituan’s native multimodal mixture-of-experts model, formally released on November 2. Making its weights available opens new options for researchers and infrastructure teams, but it does not make the model small or easy to run on an existing server.
Key points
- The Hugging Face repository lists MIT as its license and shows an update on November 9; the model’s formal release was November 2.
- The model card says 1.6 trillion total parameters and about 48 billion active per token, while the page’s file metadata reports about 1.78 trillion parameters. The discrepancy is unresolved.
- Meituan’s model card lists benchmark results, but they are vendor-reported; this search found no public leaderboard confirming an independent reproduction of those scores.
Background
LongCat-2.5 was formally released on November 2. Meituan describes it as a native multimodal model built for long-horizon agent tasks, including coding, visual understanding, visual coding and creative work. The November 9 event is specifically the release of model weights under an MIT license, not the model’s first release or the first availability of a hosted service.
That distinction matters to developers. A hosted API lets a team test a model through a service interface. Weights let teams with suitable infrastructure inspect the files, attempt fine-tuning or build their own inference service. More control also means taking on hardware, software compatibility and reliability work. Meituan’s changelog records the November 2 launch, while the Hugging Face repository shows its November 9 update; together they establish the two dates behind this story.
The facts
LongCat-2.5’s license and specifications need to be read separately: the repository declares MIT, while the parameter count is inconsistent within the Hugging Face page. The model card describes a native multimodal MoE with 1.6 trillion total parameters and roughly 48 billion active for each token. File metadata on the same page reports about 1.78 trillion parameters. The page does not explain whether this difference reflects versions, counting methods or something else, so both figures should remain visible rather than being reconciled by guesswork.
| Item | What is documented | Evidence boundary |
|---|---|---|
| Release and weights | Formal release on November 2, 2026; repository updated November 9 | Hugging Face repository timestamp and official changelog |
| License | The Hugging Face repository’s license field says MIT | A license field does not grant Meituan trademark or patent rights; check the license text for a specific use |
| Parameters | Model card: 1.6T total, about 48B active; file metadata: about 1.78T | Conflicting figures appear on the same Hugging Face page |
| Weights and deployment | The card gives weight information and NPU deployment pointers; AICoder also reports a W4A4 variant | Deployment depends on compatible hardware, software and memory; a consumer-GPU path was not verified here |
| API | Meituan’s November 2 launch coverage described online access and an API service | That does not establish that the Hugging Face files can be called through a generic public endpoint; no unverified connection recipe is included here |
The model card includes coding, agent and vision benchmarks, including Terminal Bench v2.1, BrowseComp and VideoMME. These are Meituan-reported results. The card says some competitor scores came from public sources, while others were internal or indicative. They are not the result of an independent lab running the released LongCat-2.5 weights in a matched evaluation. The card also uses context settings such as 256K, 512K and 1M for particular benchmark runs. Those settings describe individual evaluations, not a general context-window guarantee. Meituan’s materials do not state one default context limit that applies across tasks.
What others say
Independent coverage adds deployment context but has not supplied an independent scorecard. IT Home and AICoder both describe the November 9 release as an MIT-licensed weight release. AICoder also discusses W4A4 and an NPU deployment route, a reminder that the model is not a small package aimed at arbitrary graphics cards. Meituan’s published W4A4 throughput figures are company measurements; they cannot be translated directly into a speedup every user will see.
Times of AI says some competitor results in Meituan’s comparison table come from public reports and some were measured by Meituan. It reports no independent leaderboard score for LongCat-2.5 at the time of its analysis, and does not claim that its author ran the model. The defensible wording today is that Meituan reports these results, not that independent testing has confirmed them. Open weights make reproduction possible; the reproduction still has to be done.
The parameter discrepancy deserves the same care. Meituan’s model-card prose says 1.6T total parameters, while the page’s file metadata says about 1.78T. News coverage generally repeats the 1.6T specification. The metadata describes the tensors listed in the repository, while the card gives the vendor’s stated model size, but the page does not explain how the two counts relate. The gap should not be silently dropped or attributed to quantization without evidence.
Our take
The immediate audience for LongCat-2.5’s weights is a research group, inference provider or engineering team already equipped to handle a very large model. Those teams can now evaluate the files and explore adaptation under the repository’s MIT license. For an individual developer who mainly wants to check answer quality, Meituan’s hosted service or API may remain the lower-friction route. API pricing, quotas and interface details should be checked in the live platform documentation; the Hugging Face repository should not be mistaken for a confirmed public inference endpoint.
An open license does not make deployment free. The W4A4 and NPU route described by AICoder points to hardware-specific adaptation, and the Hugging Face file metadata itself describes a substantial artifact. Without testing a team’s own hardware and framework, the MIT label does not establish that the model will fit on one machine, run reliably or be economical to maintain. The release sits alongside Kimi K3.1 Pro’s open weights: more labs are putting model files and permissive licenses in developers’ hands, while reproduction quality, inference cost and usable tooling still separate a download from a deployable system.
The useful next signal is not another row of vendor scores. It is a reproducible third-party deployment report, throughput and memory measurements across hardware, and an explanation of the parameter-count difference. Teams comparing large releases can also look at our coverage of Mistral Large 4’s open weights. In each case, downloading is the start; inference support and sustained operation determine whether a model can reach a product.
How to try it
To check the license and files, open the LongCat-2.5 Hugging Face repository and inspect its MIT field, model card and current weight list. If the goal is to evaluate responses rather than host the model, start with Meituan’s previously announced online service or the platform’s current API documentation. For a private deployment, first confirm the frameworks, device support and storage requirements in the current model card, then test a small set of representative tasks. Do not treat the 256K, 512K or 1M settings in benchmark notes as a default context promise, or mistake an API example for a universal local inference command.
The confirmed facts are that the weights are public and the repository is marked MIT. Independent performance, a general context limit and the real deployment cost across hardware remain open questions. Put those items into an evaluation plan before choosing a hosted API or a self-managed deployment; that will be more useful than selecting from a vendor leaderboard alone.