Meituan Unveils 1.6 Trillion Parameter AI Model on Domestic Chips

In a landmark achievement for China's semiconductor ambitions, delivery giant Meituan has unveiled LongCat-2.0, a massive artificial intelligence model boasting 1.6 trillion parameters trained entirely on 50,000 domestic processors without using any NVIDIA GPUs. The release marks a significant milestone in China's "de-NVIDIA" strategy, demonstrating that the country's AI industry can achieve world-class performance despite US export controls on advanced semiconductors.

Meituan, best known as China's leading food delivery and local services platform, has emerged as a surprising player in the AI race. The company's research division has been quietly building AI capabilities for years, and the release of LongCat-2.0 represents the culmination of these efforts. The model is trained on a cluster of 50,000 domestic processors, a feat that required substantial engineering innovation to overcome the limitations of China-available hardware.

The achievement is particularly significant given the US government's export controls on advanced AI chips, which have restricted China's access to NVIDIA's most powerful GPUs. By demonstrating that it can train a 1.6 trillion parameter model on domestic hardware, Meituan has shown that China's AI industry is developing the capacity to operate independently of US technology. This has profound implications for the global AI landscape and the effectiveness of US export controls.

The Achievement: LongCat-2.0 by the Numbers

LongCat-2.0 represents a significant technical achievement that places it among the largest AI models ever developed. The 1.6 trillion parameter count puts it in the same league as the most advanced models from OpenAI, Anthropic, and Google, demonstrating that China's AI industry can compete at the highest levels.

The model was trained on a cluster of 50,000 domestic processors, representing one of the largest AI training clusters ever assembled in China. The training process took approximately 2.5 months, a timeframe that is comparable to what would be expected from an NVIDIA-based cluster of similar scale. This suggests that Meituan's engineering team has achieved high utilization rates and effective parallelism across the domestic hardware.

LongCat-2.0 is not Meituan's first AI model. The company previously released LongCat-1.0, a smaller model that served as a proof of concept. LongCat-2.0 represents a significant leap forward in scale and capability. The model is designed to support Meituan's core business operations, including delivery optimization, personalized recommendations, and autonomous navigation, but it also has broader applications across industries.

The achievement has been hailed by Chinese state media as a major breakthrough, with the Xinhua News Agency describing it as "a milestone in China's AI independence." The release of LongCat-2.0 comes at a time when the US has been tightening export controls on AI technology, and Meituan's success demonstrates that China's AI industry is developing the capacity to operate independently of US technology.

China's Domestic Chip Ecosystem: The 'De-NVIDIA' Strategy

LongCat-2.0 is the most visible result of China's "de-NVIDIA" strategy, a concerted effort by the Chinese government and domestic industry to reduce dependence on American semiconductor technology. The strategy encompasses multiple dimensions:

• Domestic Chip Development: Chinese companies including Huawei, SMIC, and Cambricon have been developing domestic AI chips that can compete with NVIDIA's offerings. While these chips are not yet as powerful as the most advanced NVIDIA GPUs, they are rapidly improving.
• Software Ecosystem: China is developing its own AI software stack that is compatible with domestic chips. This includes frameworks, libraries, and optimization tools that enable efficient training on domestic hardware.
• Scale and Aggregation: Chinese companies are using large clusters of less powerful domestic chips to achieve the same computational power as smaller clusters of more powerful NVIDIA chips. This requires sophisticated software and networking to manage the distributed computing.
• Government Support: The Chinese government has been providing substantial funding and policy support for domestic chip development, including the establishment of national semiconductor funds and the prioritization of domestic chip procurement for government projects.

The "de-NVIDIA" strategy is not just a response to US export controls. It is part of a broader effort by China to achieve technological self-sufficiency across critical industries. The AI sector is considered strategically important, and the Chinese government has made it a priority to develop domestic AI capabilities that are not dependent on US technology.

Meituan's achievement is a validation of this strategy. By demonstrating that a 1.6 trillion parameter model can be trained on domestic chips, Meituan has shown that the strategy can work. This is likely to accelerate investment in domestic chip development and encourage other Chinese companies to adopt domestic hardware for AI training.

However, there are still significant gaps between domestic chips and NVIDIA's most advanced offerings. The performance gap is narrowing, but domestic chips are still less powerful than the latest NVIDIA GPUs. Chinese companies are addressing this gap through scale and software optimization, but the underlying hardware limitations remain a challenge.

How They Did It: Training on 50,000 Domestic Processors

Training a 1.6 trillion parameter model on 50,000 domestic processors is an engineering feat that required significant innovation. Meituan's engineering team had to overcome several challenges to achieve efficient training on domestic hardware.

The key innovations that enabled this achievement include:

• Efficient Parallelism: The team developed sophisticated parallelism strategies that distribute the model across the 50,000 processors while minimizing communication overhead. This includes a combination of data parallelism, model parallelism, and pipeline parallelism.
• Fault Tolerance: Large clusters of processors are prone to failures, and the team implemented robust fault tolerance mechanisms that allow the training to continue despite hardware failures.
• Memory Optimization: Domestic processors have less memory than the most advanced NVIDIA GPUs, requiring careful memory management to fit the 1.6 trillion parameter model within the available memory.
• Software Optimization: The team optimized the training software stack for domestic hardware, including custom kernels and communication libraries that exploit the specific characteristics of the hardware.

The training process was not without challenges. The team reported that they had to invest significant effort in optimizing the software stack and debugging issues related to the heterogeneity of the hardware. There were also challenges related to power consumption and cooling, as the 50,000 processors consume substantial amounts of electricity and generate significant heat.

Despite these challenges, the team was able to complete the training in approximately 2.5 months. This is a credible timeframe that suggests the training was efficiently executed. The successful training of LongCat-2.0 is a testament to the skill and determination of Meituan's engineering team.

Chip Comparison: Domestic vs. NVIDIA

To understand the significance of Meituan's achievement, it is helpful to compare the domestic processors used for LongCat-2.0 with NVIDIA's most advanced GPUs. While domestic chips are improving rapidly, there are still significant differences in performance and capability.

As the comparison shows, domestic chips are still significantly less powerful than NVIDIA's most advanced GPUs. However, Chinese companies are addressing this gap through scale and software optimization. By using 50,000 domestic processors, Meituan was able to achieve the same computational power that would require far fewer NVIDIA GPUs.

The performance gap is expected to narrow in the coming years. Chinese semiconductor companies are making rapid progress, and the government is providing substantial support for domestic chip development. Meituan's achievement demonstrates that domestic chips are already capable of supporting world-class AI training, even if they are not yet competitive with NVIDIA's most advanced offerings.

China's AI Milestones: A Timeline of Progress

LongCat-2.0 is the latest in a series of milestones in China's AI development. The following timeline highlights key achievements that have built toward this moment:

This timeline reflects a decade of sustained investment and progress in China's AI and semiconductor sectors. The achievement of LongCat-2.0 is the result of years of research and development, supported by substantial government funding and policy support. The momentum is likely to continue, with further milestones expected in the coming years.

Challenges: What Still Needs to Be Solved

While LongCat-2.0 is a significant achievement, there are still challenges that need to be addressed for China's "de-NVIDIA" strategy to succeed. These challenges span hardware, software, and ecosystem development:

The performance gap is the most significant challenge. While domestic chips are improving rapidly, they are still years behind NVIDIA's most advanced offerings. Closing this gap will require sustained investment in semiconductor research and development.

The software ecosystem is another critical challenge. NVIDIA's dominance is not just about hardware—it's also about the CUDA ecosystem, which has become the industry standard for AI development. China needs to develop an independent software stack that can compete with CUDA, including frameworks, libraries, and tools that are optimized for domestic hardware.

Despite these challenges, the momentum behind China's "de-NVIDIA" strategy is strong. The government is providing substantial support, and domestic companies are investing heavily in R&D. Meituan's achievement is a sign that the strategy is working, and further progress is expected in the coming years.

Global Implications: A New Chapter in AI Competition

Meituan's achievement has significant implications for the global AI landscape. It challenges the effectiveness of US export controls and signals a new phase in the geopolitical competition over AI technology.

For US Export Controls: The success of LongCat-2.0 suggests that US export controls may not be as effective as policymakers had hoped. While the controls have imposed costs on China's AI industry, they have not prevented the development of world-class AI models. This may lead to calls for even stricter controls, or alternatively, for a shift in strategy toward collaboration and engagement.

For Global AI Competition: Meituan's achievement demonstrates that the AI race is not a two-horse race between the US and China. Other countries, including European nations and countries in Asia, are also developing AI capabilities. The global AI landscape is becoming increasingly multipolar.

For China's AI Industry: The success of LongCat-2.0 is likely to accelerate investment in domestic chip development and encourage other Chinese companies to adopt domestic hardware for AI training. This could create a virtuous cycle where more companies use domestic chips, leading to more investment, better chips, and wider adoption.

For AI Development Globally: The achievement demonstrates that world-class AI can be developed on a variety of hardware platforms. This could lead to greater diversity in the AI ecosystem, reducing the dominance of a single vendor and potentially leading to more innovation.

The release of LongCat-2.0 is a reminder that AI development is a global endeavor. While the US has been a leader in AI, other countries are making rapid progress and will play an increasingly important role in shaping the future of the technology. The challenge for policymakers is to balance the benefits of global collaboration with the imperatives of national security and economic competitiveness.

For now, Meituan's achievement stands as a testament to the resilience and ingenuity of China's AI industry. It has demonstrated that domestic hardware can support world-class AI development, and it has given China a significant boost in its pursuit of AI self-sufficiency.