Meituan's LongCat-2.0: China Trains Its Largest AI Model Ever on Homegrown Chips

Chinese tech giant Meituan has officially launched LongCat-2.0, a massive 1.6 trillion parameter open-source large language model that was developed entirely on domestic hardware — no NVIDIA chips involved. The Beijing-based food delivery company completed the project end-to-end using Chinese-made silicon, marking a significant turning point in the global AI competition.

The announcement carries particular weight at a time when China is doubling down on technological self-sufficiency, especially in the high-stakes arena of AI computing infrastructure.

What Makes LongCat-2.0 Stand Out

At its core, LongCat-2.0 is a large language model built to understand, generate, and reason across a wide spectrum of human language tasks. With 1.6 trillion parameters and a 1-million-token context window, it ranks among the largest AI models ever created. But its parameter count isn't the only headline-grabbing detail.

Meituan claims LongCat-2.0 is the first trillion-parameter model in the industry to have completed both the training and inference phases entirely on domestically produced hardware. This is a crucial distinction. DeepSeek's V4-pro, for instance, relied on Chinese chips only during inference — the comparatively lighter task of responding to user queries. LongCat-2.0, by contrast, ran the far more computationally intensive pre-training stage on home-grown chips as well.

The hardware backbone consisted of large-scale ASIC superpods — chips engineered for specific, high-demand workloads. To coordinate chip-to-chip communication at scale, Meituan deployed Huawei's Collective Communication Library, known as HCCL. The setup is structurally similar to how NVIDIA uses its own NCCL framework to manage GPU cluster coordination.

Technical Architecture and Design Choices

LongCat-2.0 employs a Mixture of Experts architecture with approximately 48 billion active parameters at any given time, despite its 1.6 trillion total. The model also features LongCat Sparse Attention, a proprietary mechanism designed to efficiently handle the full 1-million-token context length. According to Meituan, the system was built from the ground up with agentic coding capabilities in mind.

Performance Benchmarks and Global Context

In terms of raw performance, LongCat-2.0 has delivered competitive results. It outpaced Google's Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro evaluations. That said, the model still falls short of the leading Western frontier systems — OpenAI's GPT-5.5 and Anthropic's Opus 4.8 — particularly on the most advanced agentic reasoning tasks.

Reactions from the tech and investment community have been swift and pointed. Analyst TP Huang suggested the launch effectively silences lingering doubts about Huawei's Atlas-950 SuperPoDs. Hanchi Sun, a researcher at Lehigh University, described LongCat-2.0 as the first model ever trained to near-frontier performance across 50,000 domestic accelerators. Venture partner Alvin Foo offered a broader strategic read: if China can now scale frontier AI training on local chips at this level, the global compute race has become far more open than anyone had anticipated.

Analyst Yuchen Jin echoed a sentiment tied directly to US export restrictions, referencing Jensen Huang's earlier commentary: blocking NVIDIA chip exports to China may not slow AI progress there — it may simply accelerate the development of capable alternatives built on Chinese silicon.

Challenges and Remaining Gaps

Meituan has been transparent about the obstacles it faced. Memory capacity was the primary bottleneck during pre-training, as domestic accelerators currently carry less on-device memory than NVIDIA's restricted H800 chips. The broader software ecosystem around Chinese AI hardware also continues to lag behind NVIDIA's well-established developer community.

Despite these limitations, the structural message is hard to ignore. Frontier-scale model training on Chinese hardware is no longer theoretical — it is now a demonstrated reality. As a result, the performance gap between Chinese open-source models and top-tier closed Western systems may close more rapidly than most analysts had forecast.

LongCat-2.0 is now publicly available, with the full model powering Owl Alpha on OpenRouter.