Home / Technology / China's AI Leap: 1.6T AI Model Trained on Homegrown Chips
China's AI Leap: 1.6T AI Model Trained on Homegrown Chips
30 Jun
Summary
- New 1.6T MoE model boasts 1M-token context window.
- Trained entirely on Chinese ASICs, not US GPUs.
- Released under a permissive MIT license for commercial use.
Meituan has unveiled LongCat-2.0, a powerful 1.6-trillion-parameter Mixture-of-Experts (MoE) model, now publicly available on GitHub and Hugging Face. This advanced system features an unprecedented 1-million-token context window and is released under a commercially viable MIT license.
A key differentiator is its training infrastructure; LongCat-2.0 was developed entirely on a cluster of over 50,000 Chinese ASICs. This demonstrates that cutting-edge AI models can be scaled without relying on U.S.-manufactured Nvidia GPUs, potentially shifting global AI hardware dependencies.
The model offers flexible commercial access, including free processing for context-cache hits and competitive pay-as-you-go API rates. Promotional discounts further reduce operational costs, making it an attractive option for developers seeking high-performance alternatives.
LongCat-2.0 excels in autonomous software engineering tasks, achieving a score of 59.5 on the SWE-bench Pro, surpassing OpenAI's GPT-5.5. Its specialized training framework, MOPD, segregates optimization for agentic execution, reasoning, and human interaction, ensuring robust performance.
This release by Meituan, a Chinese tech giant known for its delivery app, signifies a strategic pivot into AI development. By open-sourcing frontier models, Meituan aims to become a major player in global AI infrastructure, challenging existing market dynamics.
