Meituan, a leading Chinese food‑delivery company, has stunned the tech world with the launch of LongCat 2.0, an AI model boasting an astonishing 1.6 trillion parameters. It is the largest model ever trained entirely on chips manufactured in China, and the company claims to have sidestepped U.S. export restrictions that block access to NVIDIA hardware.
This breakthrough could be a game‑changing moment for Asia’s AI ecosystem, proving that a domestic corporate titan can design, train, and deploy a massive‑scale model using only homegrown technology. It challenges the long‑standing notion that Chinese chips are limited to simple inference tasks and positions local firms to compete with global leaders in a fully independent tech environment.
Reports indicate that LongCat 2.0 leveraged a fleet of 50,000 domestic chips throughout its entire pipeline—a stark contrast to recent projects that used local components only for inference. Pre‑training, the most demanding phase where the model ingests vast amounts of data to learn language fundamentals, was achieved by interconnecting thousands of machines with Huawei’s HCCL communication framework to prevent system failures.
Industry experts were immediately captivated by the technical breakthrough, with TP Huang—reporting via the South China Morning Post—tweeting that the new SuperPoDs Atlas‑950 “eliminated any lingering doubts about Zhipu AI and DeepSeek’s ability to train large language models.” Meanwhile, researcher Hanchi Sun marveled that the model delivers performance “on the cutting‑edge frontier, trained on 50,000 domestic Chinese accelerators,” and enthusiastically declared it “the first to achieve this milestone.”
When put to the test, LongCat 2.0 proves itself a serious contender against international rivals, outperforming Google’s Gemini 3.1 Pro in programming benchmarks and intelligent assistance tasks. The company, however, acknowledges that it still trails behind titans such as OpenAI’s GPT‑5.5 and Anthropic’s Claude 4.8 Opus.
Despite the celebration, engineers admit the journey was far from smooth. Operating without U.S. hardware meant confronting a less mature GPU ecosystem, where the supporting software community lags behind NVIDIA’s robust stack. The primary technical hurdle was the limited memory capacity of local chips compared to NVIDIA’s flagship processors, creating a critical bottleneck that posed significant system‑level challenges due to both the model’s scale and the cluster’s size.
Ultimately, the team was compelled to undertake a substantial effort to build a stable, secure, and scalable infrastructure, maximizing the potential of domestic hardware and ensuring the achievement garners worldwide attention.
❓ Frequently Asked Questions (FAQ)
What is LongCat 2.0 and why is it significant?
LongCat 2.0 is an AI language model developed by Meituan that contains 1.6 trillion parameters. It is significant because it is the largest model ever trained entirely on chips manufactured in China, demonstrating that domestic hardware can support end‑to‑end training of massive‑scale AI systems.
How did Meituan train LongCat 2.0 without using NVIDIA chips?
Meituan used a fleet of 50,000 domestic chips and Huawei’s HCCL communication framework to interconnect thousands of machines. This setup enabled the entire training pipeline—data ingestion, pre‑training, and fine‑tuning—to run on homegrown hardware, bypassing U.S. export restrictions that limit access to NVIDIA GPUs.
What impact could this breakthrough have on China’s AI ecosystem?
It signals a shift toward technical self‑reliance, potentially reshaping supply‑chain dynamics and accelerating China’s push for autonomous innovation. By proving that large‑scale models can be trained on domestic chips, it challenges the notion that Chinese hardware is limited to inference tasks and positions local firms to compete with global AI leaders.
News Source: Tarreo
No comments