China has just released its largest AI model, which was trained entirely without NVIDIA chips. Meituan unveiled LongCat-2.0, an open-source large language model with 1.6 trillion parameters. The food delivery giant based in Beijing oversaw the project from start to finish with domestic hardware.

A breakthrough is now changing the global AI industry’s attitude toward China’s technological self-reliance.

What does Meituan’s LongCat-2.0 bring to the AI race?

A large language model is an AI system that is trained on massive amounts of data. These systems understand, generate, and reason about human language across many different fields. LongCat-2.0 is one of the largest, with 1.6 trillion parameters and a 1 million token context window.

The release will come as China continues its push for full self-reliance in critical computing infrastructure. In addition, Meituan said that LongCat-2.0 is the industry’s first trillion-parameter model, whose training and inference were completed entirely on domestically built hardware. The project is thus a major technical milestone.

Follow us on X to get the latest news in real time.

Introducing LongCat-2.0 🐱1.6T parameters · MoE with ~48B active · 1M contextThe full model behind Owl Alpha is @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

The key difference is significant. DeepSeek’s V4 pro used domestic chips only in the inference stage, which is a lighter task—responding to users’ questions.

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

Meituan said the cluster was built based on large-scale ASIC superpods. The chips are task-specific and customized. The company also used Huawei’s Collective Communication Library (HCCL) for large-scale coordination between chips. The solution resembles the coordination approach NVIDIA uses in NCCL for GPU clusters.

“…This reminds me of Jensen Huang’s remark on the Dwarkesh podcast: Export restrictions on NVIDIA’s GPU chips do not stop China. They only accelerate the development of AI running on Chinese chips,” analyst Yuchen Jin said on X.

Why the release of LongCat-2.0 is globally significant

LongCat-2.0 performed well across a range of benchmark tests. The model outperformed Google’s earlier Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro.

However, the model is still behind the leading international frontier models. OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 perform better on more demanding agent and reasoning tasks.

Industry observers are reacting quickly. Tech analyst TP Huang said that the release alleviates concerns related to Huawei’s Atlas-950 SuperPODs. In addition, Hanchi Sun, a researcher at Lehigh University, said that this is the first model trained to nearly state-of-the-art levels using 50,000 Chinese accelerators.

“…If China can get large-scale top-tier training working on domestic chips at this level, the computing race is more open than before,” venture capitalist Alvin Foo said.

There are still significant challenges across the broader Chinese AI ecosystem. Meituan acknowledged that its software ecosystem is still behind NVIDIA’s mature GPU community. In addition, limited memory was a major bottleneck during the pre-training stage. Domestic accelerators have less memory per device than the restricted NVIDIA H800 chip.

The broader message is about structure. Meituan’s achievement shows that state-of-the-art training is now technically possible with Chinese hardware.

As a result, the gap between Chinese open-source models and Western closed-source frontier models can narrow faster than earlier predictions.