China Responds to NVIDIA with LongCat-2.0: Only Domestic Chips, Performance Comparable to Gemini 3.1 Pro, and Open Source

Meituan, a company specialized in the sale of technology products and services, has announced LongCat-2.0, a new artificial intelligence model that stands out especially for the hardware platform used during its development. The Chinese company claims to have completed the entire training and inference process exclusively with chips developed in China, without resorting to U.S. solutions.

According to the company's statements, LongCat-2.0 has 1.6 trillion parameters and supports a context window of one million tokens. Meituan also states that performance is comparable to that of Gemini 3.1 Pro, a model presented by Google in February. The company defines the project as the first AI model with over a trillion parameters to have completed both training and inference on a cluster made up of 50,000 domestic chips. The reference to the entire development cycle represents the most relevant aspect of the announcement.

The distinction between pre-training and inference plays a central role. Inference requires fewer computing resources and consists of processing user requests by a pre-trained model. Pre-training, on the other hand, uses enormous amounts of data and requires much higher processing power. It is this phase that has so far made the most advanced GPUs decisive.

If Meituan's claims are confirmed, LongCat-2.0 would represent a concrete example of China's ability to develop large-scale AI models without relying on Nvidia hardware, a topic that has become increasingly relevant after the U.S. export restrictions imposed on chips intended for artificial intelligence. In recent years, China has accelerated investments in the development of domestic chips, aiming to build a complete ecosystem capable of supporting both hardware and software dedicated to artificial intelligence. LongCat-2.0 would thus be the first model to demonstrate the maturity of the local technological platform.

Another relevant element concerns the distribution of the project. LongCat-2.0 has been released under an open-source license, allowing the developer community to analyze the model, conduct independent checks on the declared performance, and possibly make improvements. The availability of the code also enables a direct comparison of LongCat-2.0 with the benchmarks indicated by the company and to verify the actual proximity to the performance attributed to Gemini 3.1 Pro. On the other hand, verifying the hardware infrastructure used during training is more complex, as this information comes from Meituan's own statements and cannot be confirmed or denied by external parties.

However, the announcement remains a confirmation of how the challenge between the United States and China over the future of artificial intelligence is becoming increasingly tight and highlights the need to develop domestic ecosystems to gain a real advantage over competitors. LongCat-2.0 represents only a piece of the Chinese project aiming to create a comprehensive infrastructure capable of competing with North American models.