Meituan Unveils LongCat-2.0, the First Trillion-Parameter AI Model Trained Entirely on Domestic Silicon

In a milestone that redefines the boundaries of the global artificial intelligence race, Chinese technology giant Meituan has unveiled LongCat-2.0, a frontier-class AI model boasting one trillion parameters. While the model's scale places it in the upper echelon of global AI systems, its true significance lies in its origin: LongCat-2.0 is the first model of its size to be trained entirely on domestic Chinese silicon.[1]

The announcement challenges a long-held assumption in Silicon Valley and Washington: that the sheer computational density and networking requirements of trillion-parameter models strictly require hardware from US-based Nvidia. By successfully orchestrating a massive cluster of domestic chips, Meituan has provided the most concrete evidence to date that alternative semiconductor ecosystems can support state-of-the-art AI development.[2]

Meituan, primarily known globally as a food delivery and local services behemoth, has aggressively pivoted into foundational AI research over the past three years. The LongCat-2.0 release marks its transition from an AI consumer to a primary architect, positioning the company alongside Baidu, Alibaba, and Tencent in China's domestic AI vanguard.[3]

According to the technical report released by Meituan's AI lab, LongCat-2.0 utilizes a Sparse Mixture-of-Experts (MoE) architecture. This design activates only a subset of its one trillion parameters for any given query, maximizing efficiency. In early benchmarks, the model demonstrates reasoning, coding, and multilingual capabilities roughly on par with the GPT-4 class of models, though it trails the absolute bleeding edge of recently released Western systems.[5][6]

The mechanism behind this achievement is a masterclass in overcoming hardware limitations through software ingenuity. Training a massive AI model is not merely about raw computational power, or FLOPs; it is fundamentally a networking problem. The data must flow seamlessly between tens of thousands of chips. Nvidia's primary moat has long been NVLink, a proprietary interconnect technology that allows its GPUs to share memory at blistering speeds.[4]

Domestic Chinese chips, widely believed to be the latest iteration of Huawei's Ascend series, possess formidable raw compute but historically suffer from slower interconnect bandwidth. To bypass this "memory wall," Meituan's engineers had to rethink the training architecture from the ground up.[4]

The solution came in the form of a custom distributed training framework. Meituan developed a novel 4D parallel training strategy that slices the model across the cluster in a way that minimizes the need for chips to constantly "talk" to one another. By reducing the communication overhead by an estimated 40%, they allowed the domestic chips to spend more time calculating and less time waiting for data.[5][6]

Software optimization was equally critical. The global AI ecosystem is heavily reliant on CUDA, Nvidia's proprietary software platform. Transitioning to domestic silicon requires rewriting massive codebases to run on alternative platforms like Huawei's CANN (Compute Architecture for Neural Networks). Meituan deployed a custom compiler named "Catapult" that automatically translates and optimizes tensor operations for the specific quirks of the domestic hardware.[6]

Perhaps the most daunting challenge was fault tolerance. Training a trillion-parameter model requires running a cluster of roughly 30,000 chips continuously for over 90 days. In any cluster of that size, hardware failures are a daily occurrence. If a chip fails and the system cannot recover, days of training progress—and millions of dollars in electricity—can be lost.[3][5]

Meituan's technical paper details an asynchronous checkpointing system that saves the model's state every few minutes without pausing the training process. When a domestic node fails, the system automatically isolates the dead hardware, reallocates the workload to idle backup chips, and resumes training within 45 seconds. This robust fault tolerance was the linchpin that made the three-month training run viable.[5][6]

To prove the efficacy of their stack, Meituan has open-sourced a smaller, 70-billion parameter version of LongCat-2.0, along with the inference code optimized for domestic hardware. Independent researchers have already begun testing the open-weight model, confirming that the underlying architecture is sound and highly competitive in open-source leaderboards.[3][4]

Despite the triumph, industry analysts point out the hidden costs of bypassing Nvidia. The primary trade-off is energy efficiency. Domestic silicon, manufactured on slightly older process nodes due to equipment embargoes, requires significantly more power to achieve the same computational output. Estimates suggest Meituan's training run consumed up to 45% more electricity than an equivalent run on Nvidia H100s.[2][4]

Manufacturing yields present another layer of uncertainty. While Meituan successfully acquired the 30,000 chips needed for this run, scaling to the next frontier—models with 10 trillion parameters—will require hundreds of thousands of processors. Whether China's domestic foundries can produce these chips at that volume, with high enough yields to keep costs manageable, remains an open question.[2]

Nevertheless, the geopolitical implications of LongCat-2.0 are profound. US export controls, designed to throttle China's access to advanced AI capabilities, have instead acted as a powerful catalyst for domestic self-sufficiency. By forcing Chinese tech giants to invest heavily in their own software and hardware stacks, the restrictions have inadvertently birthed a parallel, fully independent AI ecosystem.[1]

As the "silicon curtain" hardens, the global AI landscape is bifurcating. Meituan's breakthrough proves that the barrier to entry for frontier AI is no longer strictly controlled by a single Western hardware vendor. The race toward artificial general intelligence is now being run on two entirely separate tracks.[2]