NEWS → · Open weights · Meituan
LongCat-2.5 weights are out: Meituan's 1.6-trillion-parameter model is on Hugging Face under MIT
Nov 9, 2026
The weights for Meituan's LongCat-2.5 are downloadable: the official Hugging Face repository (meituan-longcat/LongCat-2.5) holds 3,562 GB of files under an MIT licence for the 1.6-trillion-parameter mixture-of-experts model, with about 48 billion parameters active per token. The model launched API-only on Nov 2, 2026 with weights marked “coming soon”; the repository was created on Nov 4 and its model card was updated again today, Nov 9. The catch for GPU renters: the card's only documented deployment route is on NPUs, not GPUs.
Image: Meituan LongCat via huggingface.co
What is in the repository, read first-hand on Hugging Face on Nov 9, 2026: a native multimodal MoE — text and image input — built for long-horizon agentic work, at the same 1.6T total / ~48B active scale as LongCat-2.0, with a 1M-token context on Meituan's API. The licence is MIT, the most permissive of the widely used ones. Meituan's own scores on the card (in-house, labelled indicative by Meituan itself) put it at 88.4 on Terminal-Bench v2.1, up from 70.8 for LongCat-2.0. Its API price, from Meituan's own platform docs checked Nov 3, is a limited-time $0.30 per million input tokens and $1.20 per million output — the hosted route exists today; the self-host route is the hard part.
The deployment watch-out, from the same card: under “Deployment” the card documents NPU platforms only, pointing to an SGLang-FluentLLM build, and its efficiency story is a W4A4 quantisation scheme built for domestic AI accelerators. No GPU serving recipe is published on the card. By our standard memory maths the BF16 weights alone are about 3,200 GB — the shipped files total 3,562 GB — so even a 4-bit build at roughly 800 GB is a multi-node job, never a single card. Treat any “runs on my GPUs” claim for this model as unverified until a GPU serving path is published and reproduced.
What it means if you rent GPUs
This is one of the largest open-weight drops of the year, and the licence is genuinely clean — MIT, no Model-as-a-Service clause of the kind Mistral attached to Large 4. But a weights drop is only useful to GPU renters when the weights run on GPUs, and today Meituan documents only an NPU path. For now the practical routes are Meituan's own API at its stated limited-time prices, or waiting for the community GPU ports that usually follow a release this size. We have added LongCat-2.5 to the model catalogue with the memory maths shown, and we will update its page when a verified GPU serving route exists.