Meituan Launches LongCat-2.0: A 1.6 Trillion Parameter Model Trained on 50,000 Domestic Computing Cards

Key Takeaways

• Massive Scale: LongCat-2.0 features a total of 1.6 trillion parameters, with an average activation of approximately 48 billion and a dynamic range of 33B to 56B.
• Domestic Infrastructure: It is the industry's first trillion-parameter model to achieve full-process training and inference on a 50,000-card domestic computing cluster.
• Ultra-Long Context: The model provides native support for a 1-million-token (1M) context window, pre-trained from the ground up.
• Coding Optimization: The architecture is specifically engineered for Agentic Coding tasks, prioritizing stable code understanding, generation, and execution.

In-Depth Analysis

Breakthrough in Domestic Computing Power and Model Scale

The release of LongCat-2.0 by the Meituan technology team signifies a major technical achievement in the utilization of domestic computing resources. By successfully training a 1.6 trillion parameter model on a cluster of 50,000 domestic cards, Meituan has demonstrated that large-scale AI development is viable outside of traditional global hardware dependencies. This model is not merely a fine-tuned version of existing architectures but was pre-trained from scratch, ensuring that the foundational weights are optimized for the specific hardware environment used.

The parameter distribution of LongCat-2.0 is particularly noteworthy. While the total parameter count reaches 1.6T, the model utilizes a dynamic activation strategy. On average, only about 48 billion parameters are activated during processing, with a dynamic range fluctuating between 33 billion and 56 billion. This approach allows the model to maintain the representational power of a trillion-parameter system while optimizing the computational efficiency required for real-world inference and training stability on a massive 50,000-card scale.

Native 1M Context and the Shift to Agentic Coding

One of the most defining features of LongCat-2.0 is its native support for a 1-million-token (1M) long context. Unlike many models that use post-training techniques to extend their context windows, LongCat-2.0 was designed with this capability from the pre-training phase. This native support is critical for the model's primary objective: Agentic Coding. In complex software development environments, models must process vast amounts of code, documentation, and execution logs simultaneously.

The architecture of LongCat-2.0 is centered around the core goal of making code-related tasks more efficient and stable. By focusing on "Agentic Coding," Meituan is positioning the model to go beyond simple code completion. The model is built to handle the full lifecycle of code interaction, including deep understanding of existing codebases, the generation of new logic, and the stable execution of code within an agentic framework. This specialization ensures that the model can act as a reliable component in automated programming workflows where long-range dependencies and high-precision execution are paramount.

Industry Impact

The launch of LongCat-2.0 has profound implications for the AI industry, particularly regarding hardware sovereignty and specialized AI applications. By proving that a 1.6T parameter model can complete its full training and inference cycle on a 50,000-card domestic cluster, Meituan has set a new benchmark for the industry's ability to scale using localized infrastructure. This reduces the barrier for large-scale AI development in regions focusing on domestic hardware ecosystems.

Furthermore, the focus on Agentic Coding and 1M context support signals a shift in the evolution of LLMs. As models move from general-purpose assistants to specialized agents capable of executing complex technical tasks, the stability and efficiency of the underlying architecture become the primary competitive advantages. LongCat-2.0’s design suggests that the future of AI in software engineering lies in models that can natively handle massive contexts and provide stable, executable outputs in real-world coding scenarios.

Frequently Asked Questions

Question: What are the specific parameter counts for LongCat-2.0?

LongCat-2.0 has a total parameter count of 1.6 trillion (1.6T). During operation, it features an average activation of approximately 48 billion (48B) parameters, with a dynamic activation range spanning from 33 billion to 56 billion (33B~56B).

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 cluster.

Question: How does LongCat-2.0 support long context windows?

LongCat-2.0 natively supports a 1-million-token (1M) ultra-long context. This support is integrated from the start of its pre-training process, rather than being added through post-processing or fine-tuning, to ensure stability in Agentic Coding tasks.