Meituan releases the trillion-parameter open-source large model LongCat-2.0, fully trained with 50,000 domestic GPUs.

```

On June 30, Meituan officially released and announced the open-source launch of its new-generation foundational large model LongCat-2.0 (Chinese name: "龙猫2.0"). Official data shows the model has a total of 1.6 trillion parameters and is the first domestic large language model to complete pre-training and inference processes using a cluster of 50,000 domestically-produced computing cards.

Compared to previous industry releases focusing on general capabilities, Meituan's emphasis this time is mainly on domestic substitution of core infrastructure, model inference cost control, and vertical optimization for agent and code scenarios.

Public information indicates that LongCat-2.0 was previously codenamed "Owl Alpha" for gray testing on platforms like OpenRouter. Technically, the model was pre-trained with 30 trillion tokens on a cluster of more than 50,000 domestic AI chips, primarily using sparse attention and dynamic activation mechanisms.

During inference, the model activates an average of roughly 4.8 billion parameters per token. This architectural design aims to reduce calls to high energy-consuming compute nodes when processing standard instructions, thereby lowering memory usage and inference costs per interaction.

On the model capability side, LongCat-2.0 is clearly positioned to focus on agent workloads. It was pre-trained from scratch, natively supports 1M super-long context, and its architecture makes it more efficient and stable for code understanding, generation and execution in real Agentic Coding tasks.

In mainstream agent and code generation evaluations, its main advantage lies in automated workflows and code comprehension. The model is currently compatible with interfaces for mainstream development tools such as Claude Code and OpenClaw. This technical focus shows Meituan is attempting to find more definitive application scenarios in long text and code generation.

With limited access to overseas high-end GPUs, Meituan started adapting to domestic computing power since 2023. The revealed 50,000-card domestic cluster proves the engineering viability of ultra-large-scale domestic hardware for trillion-parameter model training.

The core obstacle in clusters above ten-thousand cards lies in node collaboration and system error correction. Completing pre-training from scratch at this scale means its underlying operator adaptation, communication library error handling, and pipeline scheduling capabilities have reached commercial standards.

For enterprises, breaking free from dependency on single overseas hardware vendors gives them some room for cost control in future capital expenditure structure and computing power procurement.

Prior to the new model release, Meituan recently established an AI Transformation department internally. This reflects that its expectations for large models are shifting from technological exploration to substantial business process transformation.

Integrating LongCat-2.0—which focuses on agent and code capabilities—into Meituan's actual business chain mainly has two expected impact pathways:

First, structural cost reduction in internal R&D. By connecting to various development tools, AI code assistants and SQL data analysis agents derived from the model can directly reduce human and time consumption in backend development and data queries.

Second, automated business flow scheduling. Meituan’s main city logistics and local lifestyle businesses involve frequent merchant and rider interactions. The native tool invocation ability combined with 1M ultra-long context can be used to handle multi-dimensional scheduling command distribution and automated customer complaints transfer, aiming to further dilute per-transaction fulfillment costs.

Overall, the release of LongCat-2.0 is an engineering delivery based on existing domestic computing infrastructure and business needs. In a cycle where industry computing power investment keeps expanding, Meituan is trying to lower inference costs through architectural adjustments and drive internal efficiency with agentic coding as the entry point. The model’s subsequent financial performance still needs to be assessed to see if it can deliver the expected cost reduction in development and operational efficiency improvements in Meituan's high-concurrency real-world business scenarios.

Risk Warning and DisclaimerThe market involves risks and investment requires caution. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are appropriate for their particular situation. Invest accordingly at your own risk. ```