Models & LLMs

Meituan's 1.6-trillion-parameter model avoids Nvidia GPUs

Meituan has released a 1.6-trillion-parameter AI agent model trained entirely on Chinese hardware, outperforming some Western models in specific benchmarks while avoiding Nvidia GPUs.

The Decoder Β· Nov 15, 2026

What happened

  • Meituan's LongCat-2.5 model is trained on Chinese hardware
  • The model outperforms some Western models in coding benchmarks
  • Meituan's model avoids using Nvidia GPUs

Why it matters

Meituan's achievement highlights the potential of Chinese hardware in large-scale AI development, challenging Western dominance in the field. The model's performance on coding benchmarks suggests it could be a strong contender in specific tasks, while its use of native Chinese chips may indicate a shift in the global AI landscape.

The Elephant take

🐘 ιΌ‹ Meituan's model is a bold move, but it's still behind some Western models in certain areas. The fact that it uses only Chinese hardware is impressive, but it's not clear how this will impact the broader AI market.

Who should care

  • AI researchers
  • Tech companies
  • Hardware manufacturers

What to do next

  1. Monitor the model's performance in real-world applications
  2. Investigate the implications of using native Chinese hardware
  3. Compare the model with other AI agents in the market

Keep in mind

The model's performance on certain benchmarks is still slightly behind some Western models, and more testing is needed to determine its full capabilities.

Read the original reporting at The Decoder β†—