China's Meituan announced that its latest AI model was developed using domestic chips for training.

      LongCat-2.0, which boasts 1.6 trillion parameters, is the first model of this magnitude to be trained end-to-end on domestically produced silicon, according to the company, as a direct response to US export restrictions. What stands out about Meituan's new AI model is not just its size, but the infrastructure it utilized for training.

      On Tuesday, the Chinese delivery and services giant unveiled LongCat-2.0, claiming it is the first model of this scale to be fully trained on home-developed chips. This achievement directly targets the export controls imposed by Washington that limit access to advanced silicon for Chinese companies.

      The model's specifications are impressive, as LongCat-2.0 includes 1.6 trillion parameters with a context window of one million tokens. Meituan asserts that its performance is on par with Google’s Gemini 3.1 Pro, released in February. The company labels it as “the industry’s first trillion-parameter model to undergo end-to-end training and inference on a 50,000-chip domestic compute cluster.”

      The model has been open-sourced, making its weights accessible to anyone interested in exploring or utilizing them. A key aspect of this announcement is the term “end-to-end.” While many Chinese models perform inference on domestic hardware, which is the less intensive task of responding to queries after training, pre-training involves the more resource-intensive process of analyzing vast datasets to learn fundamental patterns. This is where the advanced chips play a crucial role.

      Meituan’s assertion that LongCat-2.0 was both pre-trained and executed on domestic silicon elevates the significance of the announcement beyond mere marketing. If validated, it addresses a strategic concern in China's AI industry: the capability to develop frontier-scale models without relying on Nvidia.

      Washington restricts exports of cutting-edge chips for national security reasons, prompting Beijing to invest heavily in domestic alternatives, accelerating the design and manufacturing of silicon to match the capabilities of American hardware.

      This effort has yielded a series of significant achievements. Recently, China claimed the supercomputing title without US chips, and several domestic competitors have emerged to challenge Nvidia's supremacy, including Alibaba’s T-Head unit, which is promoting its Zhenwu M890 GPU as a locally developed accelerator.

      LongCat-2.0 acts as the software counterpart in this hardware initiative—an extensive model aimed at demonstrating the viability of domestic technology at scale.

      Meituan is an unexpected leader in this movement; it is better recognized for food delivery than for advanced AI. Nevertheless, the company is among several Chinese internet giants that have aggressively pursued model development, treating it as vital infrastructure rather than a secondary endeavor.

      By open-sourcing a 1.6-trillion-parameter model, Meituan is also making a competitive move, encouraging adoption among developers and indicating confidence in the ability of the underlying chips to support this effort.

      For a company that operates one of the largest on-demand logistics services globally, the benefits of affordable, locally secured AI are tangible: operations such as routing, demand forecasting, and customer service all depend on computing power, and a model trained on domestic silicon protects that power from the next wave of export restrictions.

      Validation from the open-source community will come as it runs LongCat-2.0 against the benchmarks Meituan claims, allowing for a genuine assessment of its match with models like Gemini 3.1 Pro.

      However, the assertion regarding training hardware is more difficult for outsiders to verify directly since it relies on Meituan’s account of its infrastructure, a caveat worth considering alongside the company’s assurances.

      What remains clear is the current trajectory. The competition for AI dominance between China and the United States fundamentally revolves around chips, and each model developed without American hardware diminishes the gap that the export controls aimed to create.

      Meituan’s announcement adds another piece to the ongoing contest that Washington intended to dominate with its restrictions, while Beijing is resolute in demonstrating its capability to operate independently.