Chinese technology company Meituan has released LongCat-2.5, a new native multimodal artificial intelligence model designed for complex, long-running AI-agent tasks. The model has 1.6 trillion total parameters, with approximately 48 billion activated for each token. It combines visual understanding, advanced coding capabilities and long-horizon reasoning to perform tasks such as building software projects, creating interactive websites and analysing business operations.
The November 2 launch expands Meituan’s LongCat model family and strengthens its position in China’s increasingly competitive AI industry. The company says the model was trained entirely on domestic computing infrastructure and is compatible with agent development tools such as Claude Code, Codex and Hermes. Developers can access it through an API, while users can also try it through the LongCat website. The model’s weights have not yet been publicly released.
What Is Meituan’s LongCat-2.5 AI Model?
LongCat-2.5 is a mixture-of-experts (MoE) model developed by Meituan’s LongCat AI team. Its architecture contains approximately 1.6 trillion parameters, but only around 48 billion parameters are activated for each token during inference.
This architecture allows the model to have a large overall capacity without activating every parameter for every calculation.
The new model builds on LongCat-2.0, which Meituan introduced in June 2026. While the earlier model focused on large-scale language and agentic coding capabilities, LongCat-2.5 adds stronger native visual understanding and more advanced capabilities for completing complex, multi-step tasks.
Meituan is particularly targeting long-horizon agent workflows. These are tasks in which an AI system must plan, use tools, execute multiple steps, evaluate results and continue working until it completes a larger objective.
For example, instead of generating a short piece of code, LongCat-2.5 is designed to work through a complex software project, understand its existing codebase, identify problems and implement changes across multiple files.
Native Multimodal Capabilities Combine Vision and Text
One of the major upgrades in LongCat-2.5 is its native multimodal architecture.
The model represents visual information using multi-level discrete tokens, allowing its main architecture to process visual and textual inputs through a unified token-based system.
This approach is designed to improve how the model understands images and connects visual information with instructions and reasoning.
Meituan says these capabilities support several applications:
- Visual coding: Building websites and interfaces based on visual requirements.
- Creative design: Producing interactive visual experiences and design-oriented projects.
- 3D development: Creating three-dimensional demonstrations and digital environments.
- Visual reasoning: Understanding images and using that information to complete more complex tasks.
The company has demonstrated projects including interactive fluid simulations, 3D digital exhibitions, miniature city environments and other visual applications.
These examples illustrate the model’s focus on generating functional outputs rather than simply describing what a developer or designer should build.
LongCat-2.5 Focuses on Complex AI-Agent Coding
Meituan has invested heavily in improving the model’s ability to complete extended engineering tasks.
The training process uses large-scale reinforcement learning, with an emphasis on long sequences of interactions between an AI agent and its working environment.
The company says it has expanded its training pipeline across three important areas: data synthesis, environment construction and training across different agent frameworks.
An interesting part of this approach involves using failed AI-agent attempts as training material.
When an agent fails to complete a real-world task, the development team can convert that failure into a reproducible training example with verifiable tests. This helps the model learn from specific weaknesses rather than relying exclusively on successful demonstrations.
LongCat-2.5 is also designed to work with established agent frameworks, including Claude Code, Codex and Hermes. This gives developers more flexibility to integrate the model into existing software-development workflows.
Meituan Claims Improvements in Training and Inference Efficiency
Alongside model capabilities, Meituan has focused on making LongCat-2.5 more efficient on Chinese-made AI accelerators.
The company says the model’s training process included sequences of up to one million tokens during its mid-training stage. Training such long sequences creates significant memory and communication challenges, particularly when computing hardware has limited memory capacity.
Meituan worked on memory utilisation, communication between accelerators and parallel processing to improve efficiency.
According to its technical documentation, the optimised training system achieved an 86% increase in per-device throughput compared with a Distributed Data Parallel (DDP) baseline in the tested long-sequence supervised fine-tuning tasks.
The company has also introduced W4A4 quantisation for inference. This approach represents both model weights and activations using four-bit precision, reducing the amount of data required for calculations.
Meituan’s reported online tests showed the following improvements compared with its W8A8 INT8 baseline:
| Performance metric | Reported improvement |
|---|---|
| Input processing throughput | Around 40% |
| Token generation throughput | Around 40% |
| First-token waiting time | Reduced by around 16% |
| Time required to generate subsequent tokens | Reduced by around 20% |
These figures are based on Meituan’s own testing and have not been independently verified.
API Availability and Developer Integration
LongCat-2.5 is available through Meituan’s LongCat platform and API service.
The model’s compatibility with existing agent frameworks is particularly relevant for developers who want to experiment with different AI models without rebuilding their entire workflows.
The official change log lists support for Claude Code, Codex, Hermes and other agent harnesses.
However, the release is not an open-weight launch. Meituan has not announced publicly downloadable LongCat-2.5 model weights. Its announcement says GitHub and Hugging Face releases are expected to follow.
This distinction matters for developers. API access allows applications to use the model through Meituan’s infrastructure, while an open-weight release would allow developers to download and deploy the model themselves, subject to its licence and hardware requirements.
Meituan Wants AI to Support Its Core Businesses
LongCat-2.5 is also connected to Meituan’s broader strategy of integrating AI into its existing businesses.
The company operates major food delivery and local-services platforms in China. It sees opportunities to use AI to improve merchant operations, customer experiences and internal business efficiency.
One example is Miaoshou, an AI-powered business operations workspace developed using the LongCat model family.
The platform is designed to help local merchants analyse business performance, understand operational problems and execute tasks through authorised tools and systems.
Meituan says Miaoshou has been used by brands including Guming, Xushengji and Yihe Tang. Its applications include business analysis, store management and customer acquisition.
This strategy gives the company a practical environment in which to deploy its AI models rather than relying entirely on external API customers.
The Bigger Picture
LongCat-2.5 reflects two important developments in China’s AI industry: the push towards more capable autonomous agents and the effort to develop large models using domestic computing infrastructure.
Its native multimodal design expands the range of tasks that AI agents can attempt, while its focus on long-running coding workflows targets applications that could generate commercial value for businesses.
The infrastructure work is equally important. Training and operating large models efficiently on domestic accelerators could help Chinese AI developers reduce their dependence on foreign hardware. However, Meituan’s performance figures and benchmark comparisons remain company-reported results that require independent evaluation.
Looking Ahead
The next important development will be how LongCat-2.5 performs in independent coding and multimodal evaluations, particularly on complex tasks that require sustained reasoning and repeated tool use. Developers will also be watching for API pricing, broader availability and the promised GitHub and Hugging Face releases.
For Meituan, the larger opportunity is to turn its AI research into practical products for merchants, customers and developers. If LongCat-2.5 can reliably complete complex tasks while operating efficiently on domestic computing infrastructure, it could become an important part of the company’s wider AI strategy.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
