Meituan LongCat Open Sources LongCat-2.5 Trillion-Parameter Multimodal Model

On Nov 9, Meituan LongCat open-sourced LongCat-2.5, a 1.6T-parameter native multimodal model under the MIT license. Featuring million-token context and W4A4 inference, it targets long-horizon agent tasks, supporting commercial use and fine-tuning for code, visual coding, and office scenarios.

Back to news

Industry News AI 编辑Published Nov 17, 2026 122 views

Overview

Amidst intensifying competition in large AI models, Chinese tech giants continue to draw industry attention with open-source initiatives. On November 9, Meituan’s LongCat team officially open-sourced LongCat-2.5, a native multimodal model designed for long-horizon agent tasks. This move marks the model’s transition from internal R&D to the public community, allowing global developers, research institutions, and enterprises to freely access, use, and improve the model.

配图

LongCat-2.5 boasts impressive technical specifications. With a total parameter count of 1.6 trillion, it ranks among the industry’s top tiers. Crucially, it employs a Mixture of Experts (MoE) architecture, activating only approximately 48 billion parameters per token during inference. This design maintains vast knowledge capacity and reasoning power while significantly reducing computational overhead. Additionally, the model supports a context window of up to one million tokens, enabling it to process and comprehend massive inputs, such as complete codebases, lengthy documents, or complex visual scene sequences.

Under the permissive MIT license, LongCat-2.5 allows free use, modification, and distribution, including for commercial purposes, without additional authorization or fees. This lowers the barrier for enterprise deployment and secondary development. Defined as a “native multimodal model,” it is optimized for long-horizon agent tasks, demonstrating significant end-to-end delivery capabilities in complex codebase development, full project construction, Visual Coding, and professional office scenarios.

Notably, the LongCat team conducted deep collaborative optimization in model architecture, chip adaptation, and deployment strategies to address potential limitations in domestic computing chips regarding memory and bandwidth. This ensures LongCat-2.5 can run efficiently not only on international high-end GPU clusters but also on domestic computing infrastructure, promoting the prosperity of the domestic AI ecosystem.

插图1

插图2

Background

The open-sourcing of LongCat-2.5 is not an isolated event but the result of LongCat’s long-term technical accumulation and strategic evolution. On November 2, Meituan LongCat released the model, initially open to internal testing and specific partners. After a week of iteration and optimization, the team decided to open-source it, likely aiming to leverage the open-source community to accelerate model validation and ecosystem building.

From a technical perspective, LongCat-2.5 represents a key trend in multimodal large models: shifting from simple “vision+text” concatenation to true “native multimodal” integration. Traditional models often use external vision modules, projecting image features into the text model’s embedding space. This approach can lead to loose semantic alignment and information loss in complex visual tasks.

LongCat-2.5’s excellence in Visual Coding stems from fundamental changes to its visual representation layer. The model represents visual signals as multi-level discrete tokens, processed by the same backbone model as text tokens. Specifically, it applies residual vector quantization to native-resolution visual features, progressively adding detail. This treats visual information as a first-class citizen in the same semantic space as text, rather than a translated accessory. Visual tokens can be precomputed and cached offline, reducing online inference load and balancing distributed training. Joint training of the visual codebook and backbone model ensures semantic consistency in tasks like visual understanding and code generation.

Regarding inference efficiency, LongCat-2.5 adopts a W4A4 efficient inference scheme. This quantizes both model weights and activations to 4-bit precision. By executing calculations via native INT4 matrix multiplication units, this scheme significantly reduces memory usage and bandwidth requirements with minimal loss in model accuracy. For a 1.6 trillion-parameter model, W4A4 quantization is a key technical support for practical usability.

Furthermore, alongside the model release, LongCat built an API platform and technical blog, forming a complete ecosystem of “open-source model + API service + technical documentation.” This strategy serves both the open-source community and commercial users seeking rapid integration. LongCat-2.5’s open-sourcing marks Meituan’s shift from “self-use” to “public utility” in AI infrastructure, aiming to attract developers to explore long-horizon agent possibilities through open sharing.

插图3

Impact

The open-sourcing of LongCat-2.5 is likely to have profound impacts on the AI industry, developer communities, and market dynamics.

First, for developers, LongCat-2.5 provides a powerful, free, and commercially usable multimodal base model. Its 1.6 trillion parameters give it the capacity to compete with top international closed-source models. The million-token context window enables developers to build agents capable of handling ultra-long documents, complete codebases, or complex multi-turn conversations. The MIT license eliminates legal and commercial authorization concerns, encouraging innovation and secondary development. Developers can fine-tune LongCat-2.5 to build specialized agents for vertical fields like law, healthcare, or finance, or leverage its Visual Coding capabilities to create new creative tools.

Second, in terms of industry trends, LongCat-2.5 accelerates the adoption of “native multimodal” and “long-horizon agent” technologies. Its visual representation layer design offers a new technical reference, potentially prompting more institutions and vendors to abandon simple external vision modules in favor of deeper multimodal fusion architectures. The successful application of W4A4 quantization proves the feasibility of high-precision models under low-precision inference, helping reduce deployment costs and pushing AI applications from the cloud to the edge.

Third, for the domestic computing ecosystem, the deep collaborative optimization between LongCat’s model architecture and domestic chips is exemplary. Amidst intensifying global computing competition, ensuring large models run efficiently on domestic hardware is key to maintaining AI industry autonomy. LongCat-2.5’s strong performance on domestic chips will boost confidence among domestic computing vendors, attracting more software ecosystem adaptations and forming a positive “hardware-software-application” cycle.

Finally, from a market competition perspective, Meituan LongCat aims to build a developer ecosystem centered on its model through open-sourcing LongCat-2.5. This enhances Meituan’s technical brand influence in AI and lays the foundation for integrating stronger AI capabilities into core businesses like local life and e-commerce. The open-source strategy may also pressure international closed-source model vendors, shifting global AI competition from pure model performance to a comprehensive contest of ecosystem building, open-source contributions, and application落地.

In summary, Meituan LongCat’s open-sourcing of LongCat-2.5 is not just a technical showcase but a strategic ecosystem layout. By providing a high-performance, low-cost, and easy-to-deploy multimodal model, it is expected to accelerate the popularization and application of long-horizon agent technologies across industries, advancing AI from “usable” to “user-friendly.”