Meituan’s Longcat 2.0: China trains a massive AI model without Nvidia
Chinese tech giant Meituan has released Longcat 2.0, a large language model trained entirely without Nvidia hardware. The open-source model demonstrates that China can develop competitive AI systems despite U.S. export restrictions on advanced chips.
Longcat 2.0 uses a mixture-of-experts architecture with 1.2 trillion total parameters and 74 billion activated parameters per token. It was trained on AMD GPUs and Chinese-made accelerators, bypassing the Nvidia ecosystem that dominates most Western AI development.
How Longcat 2.0 compares to Western models
The model achieves performance on par with Meta’s Llama 3.1 70B and Qwen 2.5 72B on standard benchmarks. Meituan also says Longcat 2.0 outperforms GPT-4 on some internal tasks, though independent verification is still pending.
Key technical details:
- Training hardware: 2,048 AMD MI250 GPUs paired with 1,200 Chinese-made Ascend 910B chips.
- Training cost: Estimated at $3 million, significantly lower than comparable models from OpenAI or Google.
- Training data: 4.8 trillion tokens drawn from open Chinese and English sources.
- Context length: 128,000 tokens, matching many frontier models.
Why this matters for the global AI race
The U.S. ban on exporting Nvidia’s A100 and H100 chips to China forced companies like Meituan to find alternative hardware. Longcat 2.0 proves that Chinese firms can still train world-class models using AMD and domestic chips.
“This is a clear signal that export controls alone will not stop China’s AI progress,” said one analyst cited in the report. “They are innovating around the restrictions.”
The model is released under a permissive open-source license, allowing researchers and developers to inspect, modify, and deploy it freely. Meituan has also published detailed training logs and optimization techniques to help others replicate the process.
What Longcat 2.0 means for open-source AI
Meituan’s decision to open-source Longcat 2.0 contrasts with many Western companies that keep their largest models proprietary. The move could accelerate AI development in China and other regions with limited access to Nvidia hardware.
Implications for developers:
- Lower entry barrier: Teams can train large models without Nvidia GPUs, using cheaper or domestic alternatives.
- Performance portability: The training techniques developed by Meituan can be applied to other AMD or Chinese hardware stacks.
- Supply chain resilience: Reliance on a single GPU vendor is no longer a necessity for state-of-the-art AI.
The bottom line
Longcat 2.0 is a technical and strategic milestone. It proves that large-scale AI training is feasible without Nvidia, and that open-source collaboration can thrive under geopolitical pressure.
Meituan has already deployed Longcat 2.0 internally for tasks like customer service and logistics optimization. The company plans to release future versions that further close the gap with top Western models.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.