Meituan from China announced that its latest AI model was developed using domestic chips.

      LongCat-2.0, a model with 1.6 trillion parameters, is claimed by the company to be the first of its magnitude to be trained entirely on domestically developed silicon, serving as a direct response to US export controls. While its size is impressive, the real significance lies in the hardware it utilizes.

      On Tuesday, the Chinese delivery and services giant Meituan introduced LongCat-2.0, asserting that it is the first model of this scale to undergo complete training on locally produced chips, which is a significant achievement amid the export restrictions imposed by Washington to prevent cutting-edge silicon from reaching China.

      The specifications are substantial. LongCat-2.0 features 1.6 trillion parameters and a context window of one million tokens, with Meituan stating its performance is on par with Google’s Gemini 3.1 Pro, which was launched in February. The company labels it "the industry’s first trillion-parameter model to achieve end-to-end training and inference on a domestic compute cluster of 50,000 chips."

      The model has been made open-source, giving access to its weights for anyone interested in running or analyzing it. A key aspect is the term “end-to-end.” Many Chinese models utilize domestic hardware for inference—the relatively light task of responding to queries post-training. Pre-training, which entails the intensive computation where models learn patterns from extensive datasets, is where superior chips are most crucial. Meituan's assertion that LongCat-2.0 was both pre-trained and operated on domestic silicon elevates this announcement beyond mere marketing.

      If accurate, it directly addresses a strategic issue in China's AI sector: the potential to create top-tier models without reliance on Nvidia. Washington imposes restrictions on the export of cutting-edge chips for national security reasons, prompting Beijing to invest heavily in domestic alternatives, thereby expediting efforts to design and manufacture silicon capable of handling workloads previously managed by American technology.

      This initiative has led to a series of achievements; recently, China claimed supremacy in supercomputing without US chips, and various domestic contenders are starting to challenge Nvidia’s dominance, with Alibaba’s T-Head unit promoting its Zhenwu M890 GPU as a local solution.

      LongCat-2.0 is a software reflection of this hardware advancement, showcasing a large model intended to demonstrate that the domestic setup can operate effectively at scale. Meituan, primarily recognized for food delivery, is somewhat an unexpected leader in this endeavor. However, like other major Chinese internet companies, it has aggressively pursued model development, treating it as vital infrastructure rather than a secondary effort.

      Making a 1.6-trillion-parameter model open-source is also a strategic decision, fostering adoption among developers and demonstrating belief in the competence of the underlying chips. For a company managing one of the world's largest on-demand logistics networks, the advantages of affordable, domestically sourced AI are practical: routing, demand forecasting, and customer service all depend on computing capabilities, and a model trained on local silicon shields that computing from the next round of export restrictions.

      Independent validation will come from the open-source community, which can now evaluate LongCat-2.0 against the benchmarks provided by Meituan to see if it truly competes with a model like Gemini 3.1 Pro. However, the claim regarding training hardware is more challenging for outsiders to verify since it relies on Meituan’s description of its infrastructure, warranting caution alongside the company's assertion.

      The trajectory of this development is clear, however. The competition for AI supremacy between China and the United States has fundamentally become a competition over chips, and each model developed without American technology decreases the gap the export controls aimed to create. Meituan's announcement serves as another indication in a rivalry that Washington designed its restrictions to dominate, while Beijing is resolute in demonstrating its capability to operate independently.