China has just unveiled its largest AI model, which was trained entirely without NVIDIA chips. Meituan launched LongCat-2.0, an open-source large language model with 1.6 trillion parameters. The Beijing-based food delivery company ran the entire project end to end on its own domestic hardware.

This breakthrough is now changing how the global AI industry views China’s push for technological independence.

What Meituan’s LongCat-2.0 adds to the AI race

A large language model is an AI system trained on massive datasets. These systems can understand, generate, and reason with human language across many different subjects. LongCat-2.0 is among the largest models ever, with 1.6 trillion parameters and a context window of 1 million tokens.

The launch comes at a time when China continues to push for full independence in critical computing infrastructure. Meituan also said that LongCat-2.0 is the first model in the industry with more than a trillion parameters that was both trained and deployed on domestic hardware. This is a major technical milestone.

Follow us on X to get the latest news right away.

Introducing LongCat-2.0 🐱1.6T parameters · MoE with ~48B active · 1M contextThe full model behind Owl Alpha on @OpenRouter — now available.Built for agentic coding from the ground up:◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens◆… pic.twitter.com/zum2SdZ0Z2

— Meituan LongCat (@Meituan_LongCat) June 30, 2026

This difference matters. DeepSeek’s V4-pro used domestic chips only for inference—that is the simpler task of answering users’ questions.

LongCat-2.0, on the other hand, used its own custom hardware for both inference and the much heavier pre-training phase.

Meituan said the cluster is built with large ASIC superpods. These are chips made specifically for certain tasks. The company also used Huawei’s Collective Communication Library (HCCL) to enable coordination between chips at large scale. This setup is similar to how NVIDIA’s NCCL directs its own GPU clusters.

“…It reminds me of what Jensen Huang said in the Dwarkesh podcast: export restrictions on Nvidia GPUs won’t stop China. They actually accelerate the development of AI running on Chinese chips,” analyst Yuchen Jin said on X.

Why the launch of LongCat-2.0 is important worldwide

LongCat-2.0 showed strong performance across various benchmarks. On Terminal-Bench 2.1 and SWE-Bench Pro, the model performed better than Google’s older Gemini 3.1 Pro.

The model still falls behind the world’s most advanced systems. This includes systems such as OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 on the most difficult agent and reasoning tasks.

Reactions came quickly from the industry. Tech analyst TP Huang said the launch alleviates concerns about Huawei’s Atlas-950 SuperPoDs. And Lehigh University researcher Hanchi Sun called it the first model in the world to achieve near-frontier performance on 50,000 Chinese domestic accelerators.

“…If China can scale training at top levels on local chips to this point, the race for compute will be more open than ever,” venture partner Alvin Foo noted.

There are still major challenges in the broader Chinese AI stack. Meituan admitted that their software ecosystem still lags behind NVIDIA’s mature GPU community. Memory was also the biggest bottleneck during the pre-training phase. So domestic accelerators have less memory than NVIDIA’s restricted H800 chip.

The big signal is structural. Meituan’s success proves that frontier-level training is now technically feasible on Chinese hardware.

This allows the gap between Chinese open-source models and the best closed Western systems to close faster than many analysts had predicted recently.