Meituan, China's dominant delivery and local services platform, has quietly engineered a breakthrough in AI inference efficiency that could reshape how companies deploy artificial intelligence at scale. The company's LongCat-2.0 model, a 1.6-trillion-parameter system, activates only about 48 billion parameters per token, representing roughly a 3% activation rate while matching or exceeding the performance of leading Western models like Claude and Google's Gemini. This extreme efficiency allows Meituan to serve its smaller LongCat-Flash variant at an astonishingly low cost of $0.70 to $0.75 per million output tokens, a development that signals a fundamental shift in how test-time compute, or the computational resources used during inference, is being optimized globally.

What Is Test-Time Compute and Why Does It Matter?

Test-time compute refers to the computational resources a model uses when answering questions or generating responses, as opposed to the resources used during initial training. For years, the AI industry focused almost exclusively on scaling up training, but companies are now realizing that how efficiently a model operates during inference determines whether it can be deployed profitably at scale. Meituan's breakthrough demonstrates that you don't need to activate every parameter in a massive model to achieve excellent results; instead, intelligent routing can direct each piece of input to only the most relevant computational pathways.

This matters because inference costs directly impact whether an AI application can be commercially viable. A model that requires expensive computational resources for every query becomes prohibitively expensive to run at scale, limiting its real-world usefulness. Meituan's approach suggests that the future of competitive AI may not be about building the largest models, but about building the smartest models that know which parts of themselves to use.

How Does Meituan Achieve This Extreme Efficiency?

The core innovation behind LongCat's efficiency is an architectural technique called Mixture-of-Experts, or MoE, combined with what Meituan calls "Zero-Computation Experts." In a standard neural network, every parameter processes every input. In an MoE system, different groups of parameters, called experts, specialize in different types of tasks. A routing mechanism decides which expert should handle each piece of input.

Meituan's breakthrough is the addition of parameter-free, zero-output experts that bypass expensive computation entirely. When the routing system determines that a token, or piece of text, is straightforward to process, it can assign it to a null expert that skips the expensive Feed-Forward Network computation altogether. This means easy tasks consume almost no computational resources, while complex reasoning still gets the full computational treatment.

Steps to Understanding Meituan's Inference Optimization Strategy

• Mixture-of-Experts Architecture: Instead of using all parameters for every input, the model routes different inputs to specialized parameter groups, reducing average computational load while maintaining performance across diverse tasks.
• Zero-Computation Experts: Parameter-free routing options allow straightforward queries to bypass expensive computation entirely, reserving intensive processing for genuinely complex reasoning tasks.
• Extreme Parameter Sparsity: The LongCat-Flash model activates between 18.6 and 31.3 billion parameters on average, despite containing 560 billion total parameters, achieving roughly 95% parameter efficiency.
• Hardware-Aware Design: The architecture is optimized for domestic Chinese AI accelerators, allowing Meituan to circumvent Western hardware dependencies while maintaining state-of-the-art performance.

What Are the Real-World Performance Implications?

The practical impact is substantial. LongCat-Flash achieves throughputs of over 100 tokens per second following a 30-day training run on more than 20 trillion tokens, meaning it can generate text faster than most competing models while costing a fraction as much to operate. For enterprises, this translates to the ability to deploy AI at scale without the massive infrastructure investments that have historically been required.

The LongCat-2.0 flagship model, despite containing 1.6 trillion parameters, operates at a cost and speed profile that makes it competitive with much smaller models from Western labs. This challenges a fundamental assumption in AI development: that bigger always means better. Instead, Meituan demonstrates that architectural intelligence and efficient routing can deliver comparable results with dramatically lower operational costs.

Why Is This a Geopolitical Turning Point?

Beyond the technical achievement, Meituan's approach carries significant strategic implications. The company trained LongCat-2.0 entirely on a cluster of over 50,000 domestic Chinese AI accelerators, representing one of the largest computing clusters ever successfully mobilized outside the United States hardware ecosystem. This is particularly significant given ongoing U.S. export controls designed to limit access to advanced semiconductors.

By achieving state-of-the-art results without relying on Western silicon, Meituan has insulated its AI trajectory from future sanctions and supply chain disruptions. For global enterprises and governments seeking AI solutions that don't depend on U.S. export licensing or cloud-tracking mandates, the LongCat ecosystem presents an immediately viable alternative. This resilience ensures that Meituan can continue scaling its models independent of geopolitical pressures.

What Does This Mean for the Future of AI Inference?

Meituan's success suggests that the next frontier of AI competition may shift from raw parameter counts to inference efficiency. As models become commoditized and performance plateaus across leading systems, the ability to deliver that performance at the lowest cost becomes the differentiator. Companies that master test-time compute optimization, routing, and sparse activation will have a structural advantage in deploying AI profitably.

The LongCat ecosystem also demonstrates that innovation in AI is no longer concentrated exclusively in Silicon Valley. Meituan's architectural breakthroughs, developed using domestic hardware and trained on Chinese infrastructure, show that frontier AI research can advance through multiple independent centers of innovation. This decentralization may accelerate overall progress in the field, as different teams pursue different architectural approaches and share insights across global research communities.

For businesses evaluating AI infrastructure investments, Meituan's approach offers a clear lesson: efficiency at inference time is as important as capability. Models that waste computational resources on simple tasks, or that require expensive hardware to operate, will struggle to compete against systems designed from the ground up to optimize every computational decision. The future of AI deployment belongs to systems that are not just intelligent, but intelligently efficient.

" }