China just launched the largest AI model trained entirely without NVIDIA chips. Meituan unveiled LongCat-2.0, a large, open-source language model containing 1.6 trillion parameters. The Beijing-based food delivery giant ran the project entirely on local hardware.
This breakthrough is now reshaping how the global AI industry views China’s efforts to become self-reliant in technology.
What Meituan’s LongCat-2.0 brings to the AI race
A large language model is an artificial intelligence system that is trained on huge data sets. These systems understand, generate, and reason human language across many domains. LongCat-2.0 is among the largest ever, with 1.6 trillion parameters and a context window of 1 million tokens.
The release comes as China continues to push for complete self-reliance in critical computing infrastructure. Moreover, Meituan said that LongCat-2.0 is the industry’s first model with one trillion parameters that complements training and inference on domestic devices. As a result, the project represents a major technical milestone.
Follow us on XTo get the latest news as it happens.
The main distinction is important. Adopt DeepSeek’s V4-pro On local chips only for inference. This is the lighter task of answering user questions.
In contrast, LongCat-2.0 used home-made hardware for both inference and the more demanding pre-training phase.
Meituan He said The cluster is built around large-scale ASIC super modules. These are chipsets intended for specific workloads. Furthermore, the company used Huawei’s proprietary Collective Communication Library (HCCL) to manage large-scale inter-chip coordination. The setting reflects how NVIDIA’s NCCL formats its GPU clusters.
“…This reminds me of Jensen Huang’s point on the Dwarkesh podcast: Export controls on Nvidia GPUs won’t stop China. They will only accelerate the development of AI running on Chinese chips,” analyst Yuchen Jin He said On X.
Why is the launch of LongCat-2.0 significant globally?
LongCat-2.0 demonstrated strong performance across multiple benchmarks. It outperformed Google’s older Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro.
However, the model still follows global boundary systems. This includes OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 across the most demanding logical and reasoning tasks.
Industry observers reacted immediately. Technology analyst TB Huang He said The launch puts an end to concerns about Huawei Atlas-950 SuperPoDs. Moreover, a researcher at Lehigh University Hanshi Sun called it The first model ever trained for near-boundary performance on 50,000 Chinese domestic accelerators.
“…If China can scale up frontier training on domestic silicon at this level, the computing arms race will be more open than ever before,” project partner Alvin Fu male.
There are still major hurdles in China’s broader AI portfolio. Meituan admitted this The software ecosystem still lags behind NVIDIA’s mature GPU community. Furthermore, memory limitations were the primary bottleneck during the pre-training phase. As a result, local accelerators carry less memory per device than the blocked NVIDIA H800 chip.
The broader signal is structural. Meituan’s success proves that frontier-level training is now technically viable on Chinese hardware.
Thus, the gap between Chinese open source models and closed Western systems may be shrinking faster than recent expectations.
this post China’s LongCat-2.0 has become the largest AI model without Nvidia chips appeared first on BeInCrypto.
