---
title: "Meituan Open-Sources LongCat-2.5: A 1.6-Trillion-Parameter Multimodal Model for Long-Horizon Agents"
date: 2026-11-10
category: Foundation Models
site: NeuroAI
canonical: https://neuroai.site/a/na-auto-2026-11-10-meituan-open-sources-lon
language: en
---

# Meituan Open-Sources LongCat-2.5: A 1.6-Trillion-Parameter Multimodal Model for Long-Horizon Agents

> On 2026-11-09, Meituan open-sourced LongCat-2.5 — a native multimodal model with 1.6 trillion total parameters, up to one million tokens of context, and MIT licensing for commercial use.

## Key takeaways

- Meituan’s LongCat team released LongCat-2.5 on 2026-11-09 as an open-source model under the MIT license (IT之家, 2026-11-09).

- The model has 1.6 trillion total parameters, with 48 billion parameters activated per token, and supports up to one million tokens of context.

- It is explicitly designed for long-horizon agent tasks — including full-project code development, Visual Coding, and professional office workflows — and permits commercial inference and fine-tuning without additional authorization.

- LongCat-2.5 features a unified multimodal architecture: visual inputs are converted into multi-level discrete tokens, processed alongside text by the same transformer backbone.

- The model was co-optimized for domestic AI chips, with architectural, hardware-adaptation, and deployment-level adjustments targeting memory and bandwidth constraints.

## What happened

On 2026-11-09, Meituan announced the open-source release of LongCat-2.5 via IT之家, a major Chinese tech news outlet. The announcement confirmed that the model had already been internally released on 2026-11-02, but the formal open-sourcing occurred on that date. Unlike many prior Chinese foundation model releases — which were either closed, API-only, or restricted to research use — LongCat-2.5 carries the MIT license, granting users broad rights to deploy, modify, and commercially integrate the model without seeking further permission from Meituan. This is notable in a landscape where even ostensibly open models often impose usage caps, attribution requirements, or commercial-use restrictions in their licenses. The timing and licensing choice suggest deliberate alignment with global open-model norms — not as a symbolic gesture, but as infrastructure signaling.

The technical specifications cited are unusually precise and grounded: 1.6 trillion total parameters, million-token context, and 48 billion active parameters per token. These figures imply a highly sparse mixture-of-experts (MoE) design — consistent with industry trends toward parameter scaling without proportional compute inflation. Crucially, the report states that visual signals are represented as *multi-level discrete tokens*, then jointly processed with text in the main transformer. That is not merely multimodal fusion (e.g., CLIP-style early alignment), but true native multimodality — where vision and language share token space and attention pathways. This architecture enables end-to-end reasoning across modalities, essential for agent tasks like generating a 3D model from a sketch-and-text prompt, or navigating a UI mockup while editing backend code.

## Why it matters

LongCat-2.5 does not aim to win benchmarks. Its stated purpose — ‘long-horizon agent tasks’ — targets a different frontier: sustained, goal-directed behavior across heterogeneous inputs and extended timeframes. In practice, this means moving beyond single-turn chat or static image captioning toward agents that can manage multi-step software projects, iterate on creative assets, or coordinate cross-modal workflows in real enterprise settings. Meituan’s domain — local services, logistics, and platform operations — demands exactly such capabilities: coordinating delivery routes, interpreting storefront images, parsing restaurant menus with embedded tables and icons, or generating localized marketing copy from product photos and sales data. By releasing LongCat-2.5 openly, Meituan invites external developers to stress-test and extend these agent patterns in contexts far beyond its own stack.

The emphasis on domestic chip optimization is equally consequential. The report notes explicit co-design across model architecture, chip interface, and deployment strategy — a departure from the common practice of porting Western-optimized models onto domestic hardware via compatibility layers. This suggests Meituan invested in low-level compiler and kernel work to reduce memory bottlenecks and improve throughput. Such vertical integration is rarely visible in open releases — making LongCat-2.5 a rare public artifact of China’s pragmatic, infrastructure-first approach to AI sovereignty.

What remains unknown is how broadly the model will be adopted outside Meituan’s ecosystem. There is no public information about inference latency, quantization support, or documentation quality — all critical for production use. Nor is there third-party validation of the reported performance gains in Visual Coding or 3D generation. The source material makes no mention of training data composition, compute budget, or evaluation methodology — gaps that limit reproducibility. Still, the decision to open-source a production-grade, commercially usable, multimodal agent model under MIT terms represents a concrete step toward foundational agent infrastructure in China — not as theory, but as shipped code.

*Sources: IT之家，2026-11-09 "Meituan open-sources LongCat-2.5: a 1.6-trillion-parameter multimodal model for long-context agents" — [original report](https://www.ithome.com/1/024/556.htm)*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-auto-2026-11-10-meituan-open-sources-lon
Free to quote with attribution and a link to the original.
