Meituan’s LongCat 2.0 and the Push for a Domestic AI Stack
Analysis by Tony Fiddis, Technology and Markets Analyst
China’s artificial intelligence race has increasingly become a hardware story, and that is precisely why Meituan’s latest announcement carries so much weight. The Beijing-based super-app giant, long associated with food delivery and local services rather than frontier computing, has unveiled LongCat-2.0, an open-source large language model that the company says was trained entirely on domestically produced AI chips. If that claim holds up under independent scrutiny, it would mark a genuine milestone in China’s broader effort to build advanced AI systems without depending on Nvidia’s highest-end accelerators, a goal that has taken on new urgency amid tightening US export controls on advanced semiconductors.
The Numbers Behind LongCat-2.0
The headline figures are substantial. LongCat-2.0 is a mixture-of-experts model with 1.6 trillion total parameters, of which between 33 billion and 56 billion are dynamically activated per token, averaging around 48 billion. Meituan says pre-training consumed more than 30 trillion tokens, with the company’s technical blog citing a figure exceeding 35 trillion tokens once full training and deployment validation are included. The model also supports a native 1-million-token context window, putting it in the same scale class as DeepSeek’s V4-pro and positioning it squarely as a tool for agentic coding, repository-level software engineering, and long-document reasoning rather than casual chatbot conversation.
Crucially, Meituan claims LongCat-2.0 is the industry’s first trillion-parameter-class model to complete the full training and inference pipeline on a domestic Chinese computing cluster, specifically a 50,000-card cluster built from what the company describes only as “large-scale clusters of tens of thousands of AI ASIC superpods.” That distinguishes it from DeepSeek’s earlier approach, which reportedly used domestic chips primarily for inference while still leaning on other hardware, including Nvidia GPUs, for the more computationally demanding pre-training phase. If accurate, Meituan’s claim represents a step beyond what DeepSeek has publicly disclosed, since pre-training is far more sensitive to communication faults, memory pressure, and numerical instability at scale than inference workloads are.
The Real Story: Hardware, Not Just Benchmarks
That last point is the real story here, and it is why analysts covering the China AI chip supply chain should pay close attention. US export controls have sharply limited Chinese access to Nvidia’s most advanced AI accelerators, including recent generations of the H-series and Blackwell-class chips. As a result, Chinese technology firms have had to improvise across the entire stack. That has meant leaning harder on domestic chip designers such as Huawei’s Ascend line, Cambricon, Biren, Muxi, and Suiyuan; building much larger chip clusters to compensate for weaker individual accelerator performance; and investing heavily in software optimization and infrastructure stability to keep enormous training runs from collapsing under communication failures or loss spikes.
Meituan has been explicit that it had to build its own scalable and reliable software layer precisely because China’s domestic chip ecosystem still trails Nvidia in several important respects, including raw performance at the cutting edge, software maturity, developer tooling, and broader ecosystem support. The company says its engineering team specifically had to solve for communication faults, memory pressure, numerical stability, deterministic operators, and distributed recovery across the cluster, and that the full run completed without a rollback or unrecoverable loss spike. That is a meaningful engineering claim on its own, separate from the benchmark scores the model eventually produced.
That plausibility, however, is exactly why this announcement deserves as much caution as attention. Meituan has not identified the chip supplier or specific chip model behind the training system. Official statements and the LongCat technical blog use only the phrase “domestic chips,” and mainstream Chinese outlets covering the release, including Xinhua and Sina Tech, have repeated that same careful phrasing without naming a vendor. Huawei’s Ascend platform is the most common guess circulating in the market, largely because it is the domestic ecosystem most closely associated with large-scale AI training in China, but this remains speculation rather than confirmed fact. If the announcement is meant to demonstrate a breakthrough for China’s self-reliant AI hardware ecosystem, leaving the hardware unnamed leaves a significant hole in the story.
Benchmarks, Verification, and the Owl Alpha Reveal
There is also the matter of independent verification. Meituan reports that LongCat-2.0 scored 59.5 on SWE-bench Pro, narrowly ahead of GPT-5.5’s reported 58.6, alongside 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 73.2 on the FORTE workflow benchmark. Notably, the model had already been circulating anonymously on OpenRouter under the codename “Owl Alpha” for roughly two months before Meituan revealed its identity, during which it reportedly generated more than 10 trillion monthly tokens and ranked among the platform’s top three models by volume. That real-world adoption prior to the branded reveal lends some credibility to the performance claims, since developers were choosing the model without knowing its origin. Even so, reported benchmark performance and grand engineering claims tend to attract headlines quickly, while genuine significance depends on what happens next: independent testing by outside researchers, full access to model weights (which, as of the release, were still listed as “coming soon” on GitHub and Hugging Face), transparent technical documentation, and evidence that developers adopt the platform in meaningful, sustained numbers under the LongCat name rather than the earlier stealth identity.
A Wider Pattern for China’s AI Sector
Even accounting for these caveats, LongCat-2.0 matters because it illustrates where the Chinese AI sector is heading strategically. Rather than waiting for restored access to top-tier foreign chips, Chinese firms are being pushed toward using whatever domestic hardware is available, scaling out with larger chip clusters to offset per-chip performance gaps, compensating through software and systems engineering efficiency, and building local alternatives across the full AI stack, from silicon to model architecture to developer tooling. This approach does not necessarily close the technical gap with leading US systems from OpenAI, Anthropic, or Google, but it may narrow that gap enough for China to remain commercially and strategically competitive, particularly in enterprise software engineering and agentic coding applications where LongCat-2.0 has been specifically optimized.
The broader pressure on this sector also intersects with Meituan’s own corporate turnaround story. With Meituan shares down more than 30 percent year-to-date and CEO Wang Xing publicly acknowledging responsibility for the stock’s underperformance, the company’s aggressive pivot into domestic AI infrastructure looks as much like a strategic reinvention as a technical showcase. Whether LongCat-2.0 becomes a durable pillar of China’s domestic AI chip ecosystem, or simply a well-timed announcement, will depend on the transparency and adoption evidence that follows in the coming months.
