Meituan has released LongCat-2.0, an open source large language model containing 1.6 trillion parameters and supporting a 1 million-token context window. The model completed full process training on a computing cluster containing more than 50,000 domestic AI accelerators, making it the first trillion-parameter model to achieve that scale, and entirely avoided the use of foreign AI hardware such as those from Nvidia. Meituan claims that LongCat-2.0 showed strong performance in coding and agent-based tasks while exceeding Google’s Gemini 3.1 Pro across several benchmarks, including Terminal-Bench 2.1 and SWE-Bench Pro. Nevertheless, it acknowledged that its latest model still trails OpenAI’s GPT-5.5 and Anthropic’s Claude 4.8 Opus on broader frontier capability assessments. The company faced major engineering difficulties throughout development, noting that memory became the primary bottleneck because each domestic accelerator offered substantially less capacity than Nvidia’s H800 chip, which remains unavailable for export to China under United States rules. Engineers built additional optimisation systems to maintain stable, secure, and scalable training across the cluster.