China just debuted its biggest AI model trained entirely without NVIDIA chips. Meituan unveiled LongCat-2.0, a 1.6 trillion parameter open-source large language model. The Beijing-based food delivery giant ran the project on domestic hardware end-to-end.

The breakthrough now reshapes how the global AI industry views China’s tech self-reliance push.

What Meituan’s LongCat-2.0 Brings to the AI Race

A large language model is an AI system trained on massive datasets. These systems understand, generate, and reason in human language across many domains. LongCat-2.0 ranks among the largest ever, with 1.6 trillion parameters and a 1-million-token context window.

The release comes as China continues to push for full self-reliance in critical compute infrastructure. Furthermore, Meituan said LongCat-2.0 is the industry’s first trillion-parameter model to complete both training and inference on domestic hardware. As a result, the project marks a major technical milestone.

Follow us on X to get the latest news as it happens.

The key distinction matters. DeepSeek’s V4-pro relied on domestic chips only for inference. That is the lighter task of answering user questions.

By contrast, LongCat-2.0 used home-grown hardware for both inference and the far more demanding pre-training stage.

Meituan said the cluster was built around large-scale ASIC superpods. These are chips customized for specific workloads. Moreover, the company used Huawei’s Collective Communication Library (HCCL) to manage chip-to-chip coordination at scale. The setup mirrors how NVIDIA’s NCCL coordinates its own GPU clusters.

“…It reminds me of Jensen Huang’s point on the Dwarkesh podcast: export controls on Nvidia GPUs won’t stop China. They’ll just accelerate the development of AI that runs on Chinese chips,” analyst Yuchen Jin said on X.

Why the LongCat-2.0 Launch Matters Globally

LongCat-2.0 showed 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 trails global frontier systems. These include OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 across the most demanding agentic and reasoning tasks.

Industry observers reacted immediately. Tech analyst TP Huang said the launch puts to rest concerns about Huawei’s Atlas-950 SuperPoDs. Furthermore, Lehigh University researcher Hanchi Sun called it the first model ever trained to near-frontier performance on 50,000 Chinese domestic accelerators.

“…If China can scale frontier training on local silicon at this level, the compute arms race is wider open than ever,” venture partner Alvin Foo noted.

Significant hurdles remain across the broader Chinese AI stack. Meituan acknowledged that its software ecosystem still lags NVIDIA’s mature GPU community. Moreover, memory limits were the primary bottleneck during pre-training. As a result, domestic accelerators carry less memory per device than NVIDIA’s banned H800 chip.

The broader signal is structural. Meituan’s success proves that frontier-scale training is now technically viable on Chinese hardware.

Consequently, the gap between Chinese open-source models and the top closed Western systems may shrink faster than recent forecasts projected.

Source: BeInCrypto