Meituan Unveils LongCat-2.0: Trillion-Parameter AI Model Trained on Domestic Chips

Chinese food delivery and services giant Meituan has made a significant announcement in the artificial intelligence landscape with the unveiling of its next-generation large language model, LongCat-2.0. The company revealed on Tuesday that this new LLM is a trillion-parameter system, a scale that places it among the most advanced AI models globally. What sets LongCat-2.0 apart is its pioneering achievement: it is reportedly the first model of its size to have completed both end-to-end training and inference entirely on a 50,000-chip compute cluster powered by domestically developed Chinese processors. This development is a crucial milestone for China, which has been actively working towards greater self-reliance in advanced technology, particularly in the face of ongoing U.S. chip export restrictions. The ability to train such a massive and complex AI model using homegrown hardware demonstrates China's growing capabilities in developing its own advanced chips and reducing dependence on foreign technology. While Meituan did not disclose the specific Chinese chipmakers involved, this move highlights the increasing market share and technological advancements of domestic firms like Huawei and Enflame. Meituan claims that LongCat-2.0's performance is on par with leading international models, specifically mentioning Google's Gemini 3.1 Pro, which was released earlier in February. The model is designed to handle complex requests and boasts the capacity to process inputs of up to 1 million tokens, enabling it to work with ultra-long documents. This capability is particularly relevant for "agentic coding," where the model's architecture is optimized for efficient and reliable execution of real-world coding tasks. The LongCat team, established in 2023, is a relatively new entrant in China's competitive AI sector. Despite this, a preview version of LongCat-2.0 has already gained traction, becoming one of the three most-used models on OpenRouter, a globally popular AI marketplace. Meituan's statement also indicated that LongCat-2.0 matched or exceeded several proprietary models, including Google's Gemini, OpenAI's GPT-5.5, and Anthropic's Claude Opus, on certain coding and agent benchmarks. While the immediate integration of LongCat-2.0 into Meituan's existing business operations was not fully detailed, earlier versions of the model have been utilized to power in-app AI assistants for tasks such as recommending restaurants and booking rooms, aligning with the "agentic commerce" trend. The company also highlighted the model's versatility by demonstrating its ability to build a gaming website and even write a novel, showcasing its broad generative capabilities. This strategic move by Meituan not only advances China's AI capabilities but also signals a broader shift towards domestic innovation in critical technological sectors.