China's LongCat-2.0 Becomes the Biggest AI Model Without Nvidia Chips (tech.yahoo.com)

China has made a significant leap in AI technology with the unveiling of LongCat-2.0, a groundbreaking large language model boasting 1.6 trillion parameters, trained entirely without NVIDIA chips. Developed by Meituan, the model exemplifies China's drive for tech self-reliance, executing both training and inference on domestic hardware. This marks a pivotal moment as LongCat-2.0 is the first trillion-parameter model to utilize home-grown infrastructure for its entire lifecycle, establishing a new benchmark for the capabilities of Chinese AI systems. The technical prowess of LongCat-2.0 is evident through its 1-million-token context window and superior performance in benchmarks compared to older models like Google's Gemini 3.1 Pro. It employs innovative techniques like LongCat Sparse Attention for efficient scaling and relies on customized ASIC superpods alongside Huawei's communication technology, which mimics NVIDIA's setup. While it still trails behind leading global models such as OpenAI's GPT-5.5 in some complex tasks, the achievement signals a significant shift, suggesting that frontier-scale AI training may soon be viable on domestic Chinese hardware. This has implications for the global AI arms race, indicating that advancements in Chinese AI could occur more rapidly than previously anticipated.