On June 30th, Meituan officially released and announced the open-sourcing of its new foundational large language model, LongCat-2.0 (Chinese name "Long Mao 2.0"). Official data indicates the model boasts a total of 1.6 trillion parameters, making it the first large language model in China to have completed the full pre-training and inference pipeline using a cluster of 50,000 domestically-produced AI accelerator cards.

Unlike previous industry releases that primarily focused on general capabilities, Meituan's emphasis with this launch lies in three key areas: domestic substitution for underlying infrastructure, controlling model inference costs, and vertical optimization for agent and code generation scenarios.

Public information shows that LongCat-2.0 was previously tested under the codename "Owl Alpha" on platforms like OpenRouter. Technically, the model completed pre-training on over 30 trillion tokens using a cluster exceeding 50,000 domestic AI chips, primarily employing a sparse attention mechanism and dynamic activation.

During the inference phase, the model activates an average of approximately 48 billion parameters per token. This architectural design aims to reduce the invocation of high-energy-consumption computing nodes when processing routine instructions, thereby lowering memory usage and inference costs for each interaction.

Model Capabilities and Focus

In terms of capabilities, LongCat-2.0 is explicitly designed to focus on agent workloads. It is understood that LongCat-2.0 was pre-trained from scratch, natively supports an ultra-long context of 1 million tokens, and its architecture allows the model to perform code comprehension, generation, and execution more efficiently and reliably in real-world agentic coding tasks.

In mainstream agent and code generation benchmarks, its primary advantages point towards automated workflows and code understanding. The model is now compatible with interfaces for major development tools like Claude Code and OpenClaw. This technical focus suggests Meituan is seeking more concrete application scenarios within the realms of long-context processing and code generation.

Significance of Domestic Computing Power

Amid market conditions where access to high-end overseas GPUs is restricted, Meituan has been advancing the adaptation of domestic computing power since 2023. The disclosed 50,000-card domestic cluster validates the engineering viability of ultra-large-scale domestic hardware in actual trillion-parameter model training.

The core challenges for clusters at the ten-thousand-card level and above lie in node coordination and system error correction. Successfully completing pre-training from scratch on a cluster of this scale signifies that its underlying operator adaptation, communication library exception handling, and pipeline scheduling capabilities have reached commercial standards.

For the company, reducing reliance on a single overseas hardware supplier could provide some room for cost control in future capital expenditure structures and computing power procurement.

Strategic Business Integration

Prior to this new model release, Meituan recently established an internal AI Transformation department. This reflects a shift in its expectations for large models, moving from technological exploration towards substantive transformation of business workflows.

Integrating LongCat-2.0, with its focus on agent and code capabilities, into Meituan's actual business operations is expected to impact two main areas. First, it could lead to 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 could directly reduce the manpower and time required for backend development and data queries.

Second, it could automate workflow scheduling. Meituan's core businesses of intra-city logistics and local services involve frequent interactions with merchants and delivery riders. The 1M ultra-long context combined with native tool-calling capability could be used to handle multi-dimensional dispatch instruction distribution and automated customer complaint routing, potentially further diluting the average cost per order fulfillment.

Overall Assessment

Overall, the release of LongCat-2.0 represents an engineering delivery based on existing domestic computing infrastructure and business needs. During a period of continued industry-wide expansion in computing power investment, Meituan is attempting to reduce inference costs through architectural adjustments and using agentic coding as an entry point to drive internal efficiency gains. The model's subsequent financial performance will depend on whether it can deliver the anticipated reductions in development costs and operational efficiency improvements within Meituan's high-concurrency, real-world business environment.