Meituan Releases LongCat-2.0 AI Model Trained Entirely on Domestic Chips
LongCat-2.0 demonstrates that a large-scale AI model can be built without relying on NVIDIA chips, addressing a key bottleneck for Chinese AI development under US export restrictions.
Key Facts
- Meituan released LongCat-2.0, a 1.6-trillion-parameter MoE language model with 48 billion activated parameters per token.
- The model supports a 1-million-token context length, matching DeepSeek-V4-Pro.
- LongCat-2.0 was trained and deployed entirely on over 50,000 domestic ASIC superpods, using Huawei Collective Communication Library (HCCL) for chip-to-chip communication.
- Meituan developed LongCat Sparse Attention (LSA) to accelerate long-context processing without quality loss.
- The model's performance is comparable to Gemini 3.1 Pro across CODE AGENT, GENERAL AGENT, and FOUNDATIONAL evaluations, and it is scheduled for open release on Hugging Face.
Reporting from 1 source: GIGAZINE.
Chinese company Meituan has released LongCat-2.0, a 1.6-trillion-parameter AI model with a 1-million-token context length, trained and deployed entirely on over 50,000 domestic ASIC superpods. The model achieves performance comparable to Gemini 3.1 Pro and is scheduled for open release on Hugging Face.
Meituan announced LongCat-2.0, a large MoE language model with 1.6 trillion total parameters and approximately 48 billion activated per token. The model supports a maximum context length of 1 million tokens, matching DeepSeek-V4-Pro's specifications. Meituan developed LongCat Sparse Attention (LSA), an evolution of DeepSeek Sparse Attention, to accelerate long-context processing without quality loss.
The entire pipeline from pre-training to inference ran on over 50,000 domestic ASIC superpods. Meituan did not name the chip manufacturer but later confirmed use of Huawei Collective Communication Library (HCCL) for chip-to-chip communication. Because the domestic accelerators have less memory per device than the China-specific NVIDIA H800 (80GB), Meituan addressed the gap through parallelization strategy and memory management.
LongCat-2.0's performance is comparable to Gemini 3.1 Pro across three evaluation categories: CODE AGENT, GENERAL AGENT, and FOUNDATIONAL. The model is scheduled for open release on Hugging Face, with related code already available on GitHub.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.