On November 9, 2026, Meituan released the open weights of LongCat-2.5 on Hugging Face under the MIT license — making it one of the largest permissively licensed multimodal models currently available. The move is significant not just for its scale, but because Meituan is a food-delivery and local services company, not a dedicated AI lab.
What LongCat-2.5 Actually Is
LongCat-2.5 is a Mixture-of-Experts (MoE) language model with 1.6 trillion total parameters and approximately 48 billion active parameters per token. That architecture means it can match the quality of a dense 48B model while running inference at a fraction of the compute cost of a full 1.6T dense model. It is multimodal, handling both text and image inputs.
The model was trained on domestic Chinese AI ASICs — not NVIDIA GPUs — which is a deliberate statement about hardware independence. Its predecessor, LongCat-2.0, was also MIT-licensed and released in mid-2026, trained entirely on domestically produced accelerators. LongCat-2.5 continues that direction with improved vision understanding and reportedly better performance on agentic benchmarks.
Why MIT License Matters
The majority of large open-weight models released by Chinese companies come under custom licenses that restrict commercial use, prohibit fine-tuning for competing services, or impose geographic limitations. MIT is different: it allows commercial use, modification, redistribution, and sub-licensing without restriction, as long as the copyright notice is preserved.
At 1.6T parameters with MIT terms, LongCat-2.5 gives any organization — including Western companies — a foundation model they can build on without license negotiation. For fine-tuning shops, enterprise AI teams, and researchers who need a large MoE baseline, this is a meaningful resource that did not exist last week.
The Meituan Angle
Meituan processes hundreds of millions of delivery and local service transactions per day across China. Its AI team built LongCat as an internal tool for tasks like product search, customer support routing, and logistics optimization — the unglamorous but high-volume work that makes or breaks a consumer platform at scale.
Releasing the weights publicly is partly a talent signal (the model demonstrates what the team can build) and partly an ecosystem play. If LongCat-2.5 becomes a commonly used foundation layer, Meituan benefits from community fine-tuning, red-teaming, and extensions it did not have to build itself.
How to Get It
The weights are available directly on Hugging Face under the Meituan organization. A hosted API is available at longcat.ai for teams that want inference without running their own infrastructure. API pricing is listed as $0.30 input / $1.20 output per million tokens at an introductory rate. The weights themselves are free to download, run, and modify under MIT terms.
If you are evaluating large open-weight MoE models right now, LongCat-2.5 is worth adding to your benchmark suite. The combination of scale, vision capability, permissive licensing, and non-NVIDIA training provenance makes it one of the more interesting releases of the month — from a company most people outside China associate with takeaway noodles.