Track Hyper | Meituan open sources LongCat-Flash: Where is the strategic direction?

Author: Zhou Yuan / Wall Street News

On September 1st, Meituan officially released and open-sourced its self-developed large model LongCat-Flash-Chat. This is the first time Meituan has made a large model available as a complete product for the industry and developers.

The model adopts the industry-popular MoE (Mixture-of-Experts) architecture, with a total parameter scale of up to 560 billion (560B), but only activates 18.6 billion to 31.3 billion parameters during each inference, averaging about 27 billion, with an average activation rate of only 4.8%.

Despite such a low activation rate, Meituan claims that "the model shows significant advantages in multiple agent-related tests, while its inference speed can exceed 100 tokens/s."

Currently, the model's code and weights are fully open-sourced and licensed under the MIT License (one of the most popular and permissive open-source software licenses globally).

This move, aside from its technical significance, mainly reflects Meituan's deep considerations in its artificial intelligence strategy.

From Parameter Stacking to Engineering Balance

In the current competition of large models, the sheer scale of parameters is no longer a novel topic.

The industry has already gone through a phase of "whose model is bigger," and now it is more important to find a balance between computational constraints and deployment efficiency.

Meituan's LongCat-Flash chooses the MoE route, which activates parameters on demand through expert routing based on a massive total parameter count.

The result is that the model retains a large potential representation capability, but the actual inference cost is controlled at a level comparable to common medium to large models.

In the application process, engineering details are crucial.

Traditional MoE models often encounter issues of unstable routing and high communication costs. Meituan introduces "zero-computation experts" in the routing mechanism, allowing some tokens to quickly bypass computation, thus ensuring overall efficiency; at the same time, it increases the overlap of computation and communication through the ScMoE approach, alleviating bottlenecks during multi-node deployment.

These modifications are not flashy but touch on the real pain points of MoE implementation: how to ensure that the model runs fast and can be stably reproduced under real hardware and scheduling conditions.

Unlike some recent large models that emphasize chain reasoning and long-chain logic, LongCat-Flash is defined by Meituan as a "non-thinking foundation model."

This positioning implies Meituan's re-understanding of application scenarios.

Meituan does not attempt to prove the model's ability to achieve multi-step reasoning at the academic testing level but focuses on agent tasks: tool invocation, task orchestration, environmental interaction, and multi-round information processing in practical application layers.

This orientation is highly consistent with Meituan's business logic.

Meituan's local life services constitute a complex system involving merchant information, delivery timeliness, geographical location, inventory status, and payment rules.

A user's request often requires coordination and decision-making across multiple subsystems.

If the model can complete invocation and interaction in each link in the form of tools, it can transform AI from a mere conversational assistant into a true process engine Therefore, compared to demonstrating the "depth of thought" of the model, Meituan places greater emphasis on the model's stable execution capability, which is clearly more valuable for the business.

In Meituan's official description, LongCat-Flash's inference speed exceeds 100 tokens/s, a metric emphasized as a "significant advantage."

For industry professionals, speed has never been an isolated number; it directly reflects key variables related to deployment costs and user experience.

The MoE architecture itself has inherent challenges regarding throughput: the instability of expert routing can lead to significant differences in the time taken for different requests, and multi-card communication may hinder overall efficiency.

Meituan's ability to claim high throughput despite a large total parameter scale relies on the optimization of routing and communication. More importantly, this model can adapt to mainstream inference frameworks, including SGLang and vLLM.

This means that enterprise users do not need to significantly overhaul their deployment stack to more directly replicate the measured results.

However, from a business perspective, what enterprises are more concerned about is actually the cost per token and stability during large-scale concurrency.

A model may perform brilliantly in a single-machine environment, but if it has unstable latency under real traffic or shows a significant increase in error rates during batch requests, it will be difficult to truly become a productivity tool.

Meituan's choice is to first address scalability and throughput issues at the architectural level, and then allow developers to evaluate the cost curve through an open deployment framework.

This is the approach of "first providing a runnable baseline, then letting the market validate it," which is likely to be more meaningful in practical applications than hollow performance comparisons.

Implicit Direction of Open Source and Licensing

Unlike many domestic manufacturers that only open part of the weights or come with "non-commercial restrictions," Meituan has adopted a more thorough open-source strategy this time: weights and code are released simultaneously, and they use the MIT license.

This choice has significant implications in both legal and ecological dimensions.

From a legal perspective, the MIT license has the least restrictions, allowing free modification, distribution, and commercial use, almost placing no additional barriers for enterprise applications; this is undoubtedly a friendly signal for companies wishing to integrate the model into their own products.

From an ecological perspective, the MIT license means that Meituan is willing to treat the model as a public asset, allowing more developers to engage in secondary development and experimentation based on it. This not only accelerates the model's iteration speed but also helps Meituan make a louder voice in the fierce open-source competition.

In terms of specific operations, Meituan has chosen to release on both GitHub and Hugging Face, which represent the mainstream channels for developer communities and model distribution, ensuring that the model is quickly accessed and used.

Therefore, behind the open-source action is actually a battle initiated by Meituan for the developer ecosystem: whoever can attract more developers to experiment with their model early on is more likely to form application links and tool ecosystems in the future.

In the publicly available model card, Meituan showcased LongCat-Flash's test results across multiple benchmark dimensions: it performed outstandingly in agent-centric evaluations such as TerminalBench, τ²-Bench, AceBench, and VitaBench, while in common dimensions like general Q&A, mathematics, and code, it is basically on par with leading large models This indicates that LongCat-Flash is not aimed at completely surpassing existing mainstream models, but rather chooses a differentiated competitive path: the strengths of this model lie in multi-tool collaboration, environmental interaction, and process orchestration, which is highly consistent with the application scenarios emphasized by Meituan.

If developers wish to build a question-and-answer assistant, it may not be superior to other open-source models; however, if they want to create an intelligent agent involving multi-tool invocation, information integration, and link execution, LongCat-Flash's positioning precisely hits the market demand.

For Meituan, open-source is not just a means of external display, but also a result of integration with internal business practices.

Meituan's local life scenarios are naturally the best testing ground for intelligent agents: the delivery chain, merchant information, real-time inventory, and user interaction constitute a complex ecosystem.

If the model can stably take on the roles of tool invocation and process orchestration within this ecosystem, then Meituan's operational efficiency, user experience, and overall platform competitiveness will be enhanced.

This is also why Meituan has not focused on whether it can solve more complex logical reasoning problems, but rather on whether it can more robustly invoke tools to complete tasks.

What Meituan wants is a model that can stably complete millions of tool invocations and reduce the system error rate; clearly, Meituan believes that this has more practical value than a model that leads by a few percentage points in academic tests.

The open-sourcing of LongCat-Flash is not just an internal matter for Meituan.

In terms of the overall industry value, Meituan has provided a high-performance MoE model that can be used directly, especially as intelligent agent applications gradually become a focus of industry attention. An open-source foundation that emphasizes tool invocation and process orchestration capabilities can accelerate application exploration within the industry.

This spillover effect may manifest in two aspects: on one hand, small and medium-sized teams can quickly validate their intelligent agent products based on the model without having to build the underlying model from scratch; on the other hand, more industry scenarios (such as logistics scheduling, customer service systems, knowledge management) may also experiment with this model.

These scenarios may not be entirely the same as Meituan's local life, but they share similarities in process complexity and tool dependency.

Through the MIT open-source license, Meituan essentially provides a low-threshold infrastructure for these scenarios.

For developers, the value of LongCat-Flash lies in providing an open model that has been trained and optimized in the dimension of intelligent agents, which can be directly applied to task chains requiring tool collaboration; for enterprise users, the real test is how to embed the model into existing systems and address the compliance, monitoring, and cost issues that arise.

In this process, the most noteworthy aspect is not the accuracy of the model itself, but the stability and controllability within the process: whether it can timely degrade when invocation fails, whether it can quickly adapt when external environments change, and whether it can maintain consistent performance under high concurrency.

Only by solving these issues can the open-source model launched by Meituan truly become part of the commercial system, rather than just a technical demonstration Meituan places such great importance on the practical value of models, it is evident that the open-sourcing of LongCat-Flash is not merely a display of technical prowess, but a clear strategic statement: Meituan has chosen a path that differs from emphasizing "thinking," focusing instead on the capabilities of tools and process execution, and addressing the implementation challenges of MoE through engineering optimization.

The MIT license is characterized by being fully open-source, so this choice by Meituan not only serves its internal business but also opens up to the entire industry ecosystem.

In the future, the true value of LongCat-Flash will not lie in the scale of its parameters, but in its ability to operate stably within complex business chains, driving the application of intelligent agents from experimentation to large-scale implementation