Home / Technology / China AI Chips Train Giant Models Sans Nvidia
China AI Chips Train Giant Models Sans Nvidia
8 Jul
Summary
- LongCat-2.0 boasts 1.6 trillion parameters with a million-token context.
- Over 50,000 domestic AI accelerators trained the advanced model.
- The entire training process excluded foreign hardware like Nvidia's.
Meituan has launched LongCat-2.0, an open-source large language model featuring an impressive 1.6 trillion parameters and a million-token context window. This scale positions it competitively against other leading models. The model's training was completed on a cluster exceeding 50,000 domestic AI accelerators, a first for a trillion-parameter model.
This development signifies a major step in China's pursuit of technological self-sufficiency amidst restrictions on advanced US hardware. LongCat-2.0 not only utilized domestic chips for inference but also for the demanding pre-training stage, entirely bypassing foreign components such as Nvidia's.
The system was reportedly built on large AI ASIC superpods, employing Huawei's Collective Communication Library for processor stability. While Meituan claims strong performance in coding and agent tasks, exceeding Google's Gemini 3.1 Pro on several benchmarks, it acknowledges trailing OpenAI's GPT-5.5 and Anthropic's Claude 4.8 Opus.
Significant engineering challenges were overcome, particularly concerning memory limitations due to the lower capacity of domestic accelerators compared to unavailable Nvidia H800 chips. Though external verification is pending, the release demonstrates a clear intent to reduce Nvidia dependency in large-scale AI training.
