Meituan trained a model with 1.6 trillion parameters without using an Nvidia GPU
Mikhail T. (Sh0ny)
5 июля 20262 min read
In short
The Chinese company Meituan has released the LongCat-2.0 foundation model, which was fully trained and deployed on domestic AI ASIC supercomputers. The Mixture-of-Experts architecture model delivers strong results, but the biggest surprise is the absence of Nvidia graphics cards in the training process.
The Chinese technology company Meituan has unveiled LongCat-2.0—a foundational language model with 1.6 trillion parameters. The main feature of this release lies not in the characteristics of the neural network itself, but in the infrastructure: the model was trained and deployed on domestic supercomputers equipped with AI ASICs, without using Nvidia GPUs.
Architecture and Performance
The model is built on a Mixture-of-Experts (MoE) architecture with a total of 1.6 trillion parameters and approximately 48 billion parameters activated per token. Prior to its official announcement, LongCat-2.0 was tested on the OpenRouter platform under the codename Owl Alpha, where it ranked among the top three in terms of usage volume. In Claude Code agent scenarios, the model ranked second in the world, trailing only Claude Opus 4.8.
According to estimates by the technical community, LongCat-2.0’s agent capabilities are on par with those of Claude Opus 4.6. In programming tasks, the model slightly outperforms GLM-5.1 but falls short of GLM-5.2.
Training on Domestic Chips
A key factor is that the entire cycle—from training to inference—was built on a domestic computing cluster. According to Meituan, pre-training was conducted on more than 35 trillion tokens without any critical failures or rollbacks.
Previously, China’s achievements in the field of domestic chips were limited to narrow tasks, such as inference or fine-tuning of pre-trained models. In the case of LongCat-2.0, however, we are talking about the full cycle of creating a trillion-parameter model from scratch.
Hardware Details
Meituan has not officially disclosed the exact chip models or their quantities. The company’s materials mention only “domestic AI chips” and “AI ASIC superpods.”
However, Chinese media and the tech community estimate that approximately 50,000 accelerators were used for training. Based on indirect evidence, such as support for 200 Gbps RDMA and 64 GB of HBM per die, experts conclude that equipment of the Huawei Ascend 910C class was used. There has been no official confirmation from Meituan or Huawei on this matter yet.