tech
Behind the Delivery Riders: Meituan Quietly Built a Trillion-Parameter AI Model
Want to read in a language you're more familiar with?
TL;DR
- Meituan trained LongCat-2.0, a 1.6 trillion-parameter AI model, on 50,000 domestic AI accelerator cards.
- This is the world's first trillion-parameter model trained and inferenced entirely on China-made chips.
- The model's development leverages the vast, real-time logistics data generated by Meituan's food delivery platform.
- LongCat-2.0's architecture is optimized for handling complex, real-world scenarios and a large context window.
- The project signifies a breakthrough for China's semiconductor ecosystem, proving domestic chips can support frontier-scale AI training.
- The foundation model powers various internal AI agents for Meituan, including code generation and recommendation systems.
- This development aligns with a trend of Chinese internet giants using proprietary data to grow AI capabilities.