Meituan Launches LongCat-2.0: A Trillion-Parameter Model Trained on 50,000-Card Domestic Computing Clusters

Key Takeaways

• Massive Scale on Domestic Hardware: LongCat-2.0 is a trillion-parameter model (1.6T total) trained and deployed entirely on a 50,000-card domestic computing cluster.
• Efficient Activation Architecture: Despite its 1.6T total parameters, the model maintains efficiency with an average activation of 48B and a dynamic range between 33B and 56B.
• Ultra-Long Context Support: The model natively supports a 1M (one million) token context window, allowing for the processing of massive codebases or documents.
• Specialized for Agentic Coding: The architecture is purpose-built to handle complex coding tasks, including understanding, generation, and execution within autonomous agent frameworks.
• Pre-trained from Scratch: Unlike models that rely on fine-tuning existing weights, LongCat-2.0 underwent a full pre-training process to ensure architectural integrity.

In-Depth Analysis

Breakthrough in Domestic Computing Infrastructure

The release of LongCat-2.0 by the Meituan technology team signifies a pivotal moment for the domestic AI industry. The most striking aspect of this development is the successful utilization of a 50,000-card domestic computing cluster for the full lifecycle of a trillion-parameter model. Training a model with 1.6 trillion parameters requires immense computational power, high-speed interconnects, and sophisticated software-hardware orchestration. By completing both the training and inference phases on domestic hardware, Meituan has demonstrated that large-scale AI development is increasingly viable within the domestic ecosystem, reducing reliance on external hardware providers for top-tier model training.

This achievement is not merely about the number of cards but the stability of the training process. Managing 50,000 cards simultaneously involves overcoming significant challenges in parallelization, fault tolerance, and energy efficiency. LongCat-2.0 serves as a proof of concept that domestic clusters can now support the most demanding workloads in the field of artificial intelligence, specifically the "from scratch" pre-training of models that exceed the trillion-parameter threshold.

Architectural Efficiency and Dynamic Activation

While the total parameter count of LongCat-2.0 stands at a staggering 1.6T, the model utilizes a sophisticated activation strategy to manage computational costs. The data reveals an average activation of approximately 48B parameters, with a dynamic range spanning from 33B to 56B. This suggests a highly optimized architecture—likely utilizing techniques similar to Mixture-of-Experts (MoE)—where only a fraction of the total parameters are engaged for any given task.

This dynamic range is crucial for balancing performance and inference speed. By activating between 33B and 56B parameters, the model can provide the reasoning depth of a trillion-parameter system while maintaining the operational efficiency of a much smaller model. This balance is particularly important for real-time applications and complex coding environments where latency and resource consumption are critical factors. The ability to scale activation based on the complexity of the input allows LongCat-2.0 to remain stable and efficient across diverse tasks.

Native 1M Context and the Future of Agentic Coding

LongCat-2.0 is designed with a core focus on "Agentic Coding." This refers to a paradigm where AI does not just suggest snippets of code but acts as an agent capable of understanding entire project structures, generating comprehensive solutions, and executing code to verify results. To support this, Meituan has implemented native support for a 1M ultra-long context window.

In the context of software engineering, a 1M token window allows the model to ingest and "remember" the entirety of a large-scale codebase, including documentation, library dependencies, and historical versioning data. This eliminates the limitations of smaller context windows that often force models to lose track of global variables or architectural patterns in large projects. By focusing on code understanding, generation, and execution, LongCat-2.0 aims to provide a more stable and reliable tool for developers, moving beyond simple completion toward full-scale autonomous coding assistance.

Industry Impact

The launch of LongCat-2.0 has profound implications for the AI industry, particularly in the realm of hardware autonomy and specialized model development. First, it validates the capability of domestic computing clusters to handle the world's most complex AI training tasks, which may accelerate the adoption of domestic hardware across other technology sectors.

Second, the shift toward "Agentic Coding" models suggests a trend where general-purpose LLMs are being refined into highly specialized tools for high-value industries like software development. By optimizing for the full lifecycle of code—understanding through execution—Meituan is setting a new standard for what AI-assisted engineering looks like. This could lead to a significant increase in developer productivity and a shift in how software is architected and maintained in the future.

Frequently Asked Questions

Question: What are the total and active parameter counts for LongCat-2.0?

LongCat-2.0 features a total of 1.6 trillion (1.6T) parameters. However, it operates with an average activation of approximately 48 billion (48B) parameters, with a dynamic activation range between 33B and 56B depending on the task requirements.

Question: What hardware was used to train LongCat-2.0?

The model was trained and is run for inference on a domestic computing cluster consisting of 50,000 cards. It is the first trillion-parameter model to complete the full training and inference process on such a domestic cluster.

Question: What is the primary application focus of this model?

LongCat-2.0 is specifically designed for "Agentic Coding" tasks. Its architecture is optimized for the efficient and stable understanding, generation, and execution of code, supported by a native 1 million (1M) token ultra-long context window.