Meituan's LongCat-2.0: China's Largest AI Model Built Without a Single Nvidia Chip
Meituan has unveiled LongCat-2.0, a 1.6 trillion parameter open-source AI model trained entirely on domestic Chinese hardware without any Nvidia chips. The launch signals a major leap in China's AI self-reliance and reshapes global assumptions about compute independence.
Chinese tech company Meituan has officially launched LongCat-2.0, a massive open-source large language model boasting 1.6 trillion parameters — and not a single Nvidia GPU was used in the process. The Beijing-based food delivery giant completed both training and inference entirely on domestic hardware, marking what many experts are calling a pivotal moment in China's technological self-sufficiency journey.
LongCat-2.0 is not just large by any standard metric. With 1.6 trillion parameters, a mixture-of-experts (MoE) architecture featuring roughly 48 billion active parameters, and a one-million-token context window, it ranks among the most powerful models ever released publicly. Designed specifically for agentic coding tasks, the model introduces a proprietary LongCat Sparse Attention (LSA) mechanism that enables efficient scaling across its full context length.
What truly sets LongCat-2.0 apart from previous Chinese AI milestones is the scope of its hardware independence. Earlier efforts, such as DeepSeek's V4-Pro, relied on domestic chips only for inference — the comparatively lighter task of generating responses to user queries. LongCat-2.0, however, used Chinese-made accelerators throughout the far more computationally intensive pre-training phase as well. Meituan built its compute cluster around large-scale ASIC superpods, chips optimized for specific workloads, and leveraged Huawei's Collective Communication Library (HCCL) to coordinate communication across chips at scale — a function analogous to what Nvidia's NCCL does for its own GPU clusters.
The model's benchmark results are noteworthy. LongCat-2.0 outperformed Google's Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro evaluations. That said, it still falls short of the current global frontier, including OpenAI's GPT-5.5 and Anthropic's Opus 4.8, particularly on the most demanding reasoning and agentic tasks.
Industry voices responded swiftly. Tech analyst TP Huang stated that the launch effectively silences lingering doubts about the readiness of Huawei's Atlas-950 SuperPoDs for frontier-scale workloads. Lehigh University researcher Hanchi Sun described LongCat-2.0 as the first model ever trained to near-frontier performance on a cluster of 50,000 domestic Chinese accelerators. Venture partner Alvin Foo put it plainly: if China can achieve this level of training on local silicon, the global compute race is far more open than most analysts recently assumed.
Analyst Yuchen Jin echoed sentiments shared widely in the AI community, pointing out that export controls on Nvidia hardware appear to be accelerating — rather than hindering — the development of capable Chinese AI infrastructure.
Challenges, however, remain real. Meituan openly acknowledged that its software ecosystem has not yet reached the maturity of Nvidia's well-established GPU development community. Memory capacity also proved to be a key constraint during pre-training; domestic accelerators currently carry less on-device memory than Nvidia's now-banned H800 chips.
Despite these limitations, the structural implication of LongCat-2.0's release is hard to ignore. Frontier-scale AI training on Chinese hardware is no longer theoretical — it is a demonstrated reality. The gap between China's open-source models and the top closed Western systems could narrow considerably faster than industry forecasts had anticipated.
