# Meituan Releases LongCat 2.5, an MIT-Licensed MoE

> The Chinese delivery giant expands its open-model push with a multimodal, agentic mixture-of-experts system.

Published by The Open Weights on Nov 4, 2026. Canonical: https://theopenweights.com/news/longcat-2-5-k5ih

## Key facts

- Company: Meituan
- Model: LongCat 2.5
- Version: 2.5
- Category: Text / LLM
- Modalities: Text / LLM, Any-to-Any, Vision-Language
- License: MIT (Open weights, commercial use allowed)
- Context window: 524K tokens
- Architecture: LongcatCausalLM
- Precision: BF16
- Weights size: 3562 GB
- Significance: notable
- Published: 2026-11-04
- Last verified: 2026-11-17
- Hugging Face: https://huggingface.co/meituan-longcat/LongCat-2.5
- Canonical URL: https://theopenweights.com/news/longcat-2-5-k5ih

Meituan has published LongCat 2.5 on Hugging Face, continuing the company's steady move into open-weight AI. The model is a mixture-of-experts (MoE) system released under the permissive MIT license, and it is positioned as multimodal — handling text alongside vision inputs — with an emphasis on agentic tasks ([model card](https://huggingface.co/meituan-longcat/LongCat-2.5)).

The MoE design is notable because it lets a model carry a large total parameter count while activating only a fraction per token, which keeps inference cost manageable relative to a dense model of equivalent capacity. Meituan has not published a specific parameter count or context-length figure in this record, so the finer details of scale remain to be confirmed from the repository.

## Why it matters

Meituan is better known in the West as a food-delivery and local-services platform than as an AI lab, so its continued investment in frontier-style models is a reminder of how broadly the open-weights ecosystem has spread beyond the usual research houses. A few things stand out:

- **MIT licensing**, which permits commercial use and modification with minimal restrictions.
- **Multimodal and agentic framing**, aligning the model with the current wave of vision-language and tool-using systems.
- **An MoE architecture**, signaling an efficiency-minded approach to scaling.

For developers, the practical value will hinge on details the release will need to make concrete: how the vision stack performs, what context window it supports, and how its agentic behavior holds up against established open models. Those questions are best answered by testing against the published weights, now available directly from Meituan's Hugging Face page.

## Get the model

- [Hugging Face](https://huggingface.co/meituan-longcat/LongCat-2.5)

## Sources

- [meituan-longcat/LongCat-2.5](https://huggingface.co/meituan-longcat/LongCat-2.5) — Hugging Face, Nov 4, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Meituan Releases LongCat 2.5, an MIT-Licensed MoE", The Open Weights, Nov 4, 2026, https://theopenweights.com/news/longcat-2-5-k5ih