When you think of Meituan, you probably think of food delivery. Often compared to DoorDash, the Chinese tech giant is famous for its massive network of couriers. But this week, Meituan made a massive flex in a completely different arena: artificial intelligence.
On Tuesday, the company announced the open-source release of LongCat-2.0, a next-generation large language model. What makes this release groundbreaking isn't just its size, but how it was built. Meituan claims LongCat-2.0 is the world's first trillion-parameter AI system trained and run entirely on a massive cluster of domestically produced, Chinese-made processors.
Here is a breakdown of why this release is sending shockwaves through the global AI and semiconductor industries.
The LongCat-2.0 Milestone
Meituan might be a late arrival to China’s fiercely competitive AI sector—squaring off against heavyweights like DeepSeek, Alibaba, and ByteDance—but they are catching up fast.
The LongCat team, which was only founded in 2023, built this new iteration from the ground up to handle massive, complex workloads.
Massive Context Window: LongCat-2.0 can process inputs of up to 1 million tokens, allowing it to seamlessly analyze ultra-long documents or entire software codebases in a single prompt.
Top-Tier Performance: Meituan claims the model matches or exceeds leading proprietary systems, including Google's Gemini, OpenAI's GPT-5.5, and Anthropic's Claude Opus on key coding and agent benchmarks.
Real-World Application: A preview version of the model was so effective at "agentic coding" (AI writing and executing code autonomously) that it quickly became one of the three most-used models on OpenRouter, a globally popular AI marketplace.
The Hardware Story: Breaking the Silicon Ceiling
The most significant aspect of LongCat-2.0 isn't its software architecture; it is the hardware running underneath it.
Since 2022, Washington has imposed strict export controls, effectively cutting China off from the most advanced AI chips produced by U.S. companies like Nvidia. In response, Chinese chipmakers like Huawei and Enflame have been racing to fill the void, securing supply deals directly with domestic AI developers.
Training an AI model with over a trillion parameters is an incredibly computationally intensive process that usually relies on American silicon. By successfully training LongCat-2.0 on a massive 50,000-chip cluster of purely domestic hardware, Meituan has proven that China's AI industry is finding viable ways around U.S. export restrictions to achieve hardware self-sufficiency.
Beyond Food Delivery
Why is a food delivery app building trillion-parameter AI models? The answer lies in "agentic commerce."
Amid weak consumer sentiment and thinning profit margins in its core business, Meituan is using AI to build highly capable in-app assistants and diversify its revenue. These aren't just simple chatbots; they are autonomous agents capable of recommending restaurants, booking hotel rooms, and completing multi-step transactions on a user's behalf. LongCat's official WeChat account even highlighted the model's ability to build a gaming website or write a novel from scratch.
By open-sourcing the technology and proving the viability of domestic hardware, Meituan isn't just upgrading its app—it is positioning itself as a central pillar in China's rapidly evolving AI ecosystem.