Chinese technology giant Meituan open-sourced its latest flagship artificial intelligence model, proving that near-frontier large language models can be pre-trained entirely on domestic hardware. The Beijing-based on-demand service and food delivery company released LongCat-2.0, a highly advanced model boasting 1.6 trillion parameters and a context window of 1 million tokens. The scale and performance of the new model place it on par with Google’s Gemini 3.1 Pro, which launched earlier this year, and DeepSeek’s V4-pro, which debuted in April.
The launch represents a defining milestone for global technology infrastructure. LongCat-2.0 is the industry’s first model of this scale to complete its entire development cycle—including pre-training and inference—on a domestic computing cluster consisting of 50,000 Chinese-manufactured computer chips. While other prominent Chinese AI groups have used local hardware to run inference, they almost universally relied on restricted U.S. chips, such as Nvidia’s A100 or H100 processors, to handle the far more demanding pre-training phase. Meituan’s success shows that Chinese firms can build, optimize, and scale trillion-parameter models without relying on Western semiconductor supply chains.
The Technical Architecture of LongCat-2.0
Building an AI model with 1.6 trillion parameters is a massive technical challenge, as standard dense architectures of this size require an extraordinary amount of computing power to process information. To overcome this limitation, Meituan designed LongCat-2.0 using a Mixture-of-Experts (MoE) architecture. Instead of activating all 1.6 trillion parameters for every single query, the model dynamically routes tasks to specific specialized sub-networks. This approach allows the system to activate only 33 billion to 56 billion parameters per token, drastically reducing the electrical and processing load on the underlying server cluster.
Complementing this efficient architecture is a massive context window of 1 million tokens. This feature allows the model to process, retain, and reason through an entire codebase, a long technical manual, or hundreds of documents in a single query. On standardized evaluation benchmarks, LongCat-2.0 posted highly competitive results. The model scored 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, which are evaluation suites specifically designed to test how well AI agents can handle real-world software engineering and programming challenges. Meituan has made the model weights and associated code publicly available on Hugging Face under the meituan-longcat organization, targeting developers who build agentic coding applications.
The Hardware Breakthrough: Scaling 50,000 Domestic ASICs
The hardware story behind LongCat-2.0 is just as significant as its software capabilities. To train the model, Meituan’s engineering team assembled a cluster of 50,000 Application-Specific Integrated Circuits (ASICs) customized for artificial intelligence workloads. By relying entirely on domestic ASICs rather than general-purpose graphics processing units (GPUs) from Nvidia or AMD, Meituan showed that alternative silicon platforms are mature enough to handle frontier-level model development.
While Meituan did not explicitly name its semiconductor supplier, the company confirmed that it utilized the Huawei Collective Communication Library (HCCL) to maintain stability across the massive hardware array. Heterogeneous clusters of this scale are notoriously difficult to coordinate, as slight delays in data transfer between chips can cause the entire training pipeline to stall. By utilizing specialized communication libraries, Meituan optimized data transfer speeds and kept the 50,000-chip cluster running continuously during the pre-training phase. This accomplishment demonstrates that software optimization and clever clustering can compensate for individual hardware performance gaps.
The Intense Demands of End-to-End Pre-Training
To understand why Meituan’s achievement is unique, one must distinguish between model training and model inference. Inference is the process where a pre-trained model generates answers to user queries. While inference requires significant power at scale, it is relatively light on computational complexity, allowing companies like Zhipu and DeepSeek to run their models on domestic chips.
Pre-training, however, is a far more intensive process. During this phase, the raw model must digest petabytes of text, code, and multimodal data to learn basic language patterns, logical relationships, and factual knowledge. Pre-training requires continuous, high-bandwidth communication between thousands of chips for months at a time. Because even a single chip failure can corrupt the training run, the physical stability of the hardware and the network interconnects must be flawless. By completing this end-to-end process on domestic silicon, Meituan has proved that alternative hardware platforms are capable of supporting the most demanding phases of AI development.
Leveraging the Huawei Collective Communication Library (HCCL)
When linking 50,000 chips together into a single cohesive supercomputer, the biggest bottleneck is not the raw speed of the individual processors, but the communication network that connects them. If chips must wait for data to transfer from neighboring nodes, they sit idle, wasting valuable computing power.
The Huawei Collective Communication Library solves this problem by providing highly optimized communication algorithms tailored for local ASIC architectures. It manages the flow of data across the superpod cluster, minimizing latency and preventing data bottlenecks during complex matrix multiplications. By utilizing this library, Meituan’s engineers managed to maintain a high level of hardware utilization throughout the training run. This software-driven optimization proves that Chinese developers can squeeze maximum performance out of alternative hardware, closing the operational gap with standard Western systems.
Geopolitical Context: How US Tech Curbs Backfired
The success of the LongCat-2.0 model comes amid an ongoing technological rivalry between the United States and China. Since late 2022, the United States has systematically tightened its export controls on advanced semiconductor technology, aiming to restrict China’s access to the cutting-edge chips designed by industry leader Nvidia. These measures, implemented on national security grounds, were intended to slow down China’s development of military applications and frontier artificial intelligence models.
However, a growing number of technology analysts suggest that these defensive regulatory measures have inadvertently backfired. Rather than stopping Chinese AI development, the restrictions have forced Chinese conglomerates to stop relying on U.S. suppliers and accelerate their investments in domestic semiconductor design and manufacturing. Meituan’s AI research team began exploring the use of domestic chips in 2023, and their latest release demonstrates that the country has built a highly viable, independent hardware ecosystem. This shift of capital and research talent toward domestic foundries has catalyzed the growth of local chip champions, creating a self-sustaining cycle of innovation that operates entirely outside of U.S. control.
The Failure of the Smuggling and Grey-Market Route
Following the imposition of U.S. export controls, a thriving grey market emerged in Asia, with brokers smuggling limited quantities of Nvidia chips into mainland China. While these smuggled chips allowed smaller research labs to continue their work, they did not offer a sustainable path to building national-level AI infrastructure.
You cannot scale a trillion-parameter training pipeline on smuggled hardware. Training models at this scale requires a highly uniform, perfectly integrated data center environment with standardized power delivery, specialized cooling, and unified networking interfaces. Acquiring 50,000 high-end GPUs through underground channels is logistically impossible, and mixing different generations of smuggled hardware creates immense engineering friction. By building and training LongCat-2.0 on a unified, domestic ASIC cluster, Meituan has demonstrated that the only viable path to long-term technological independence lies in sovereign domestic manufacturing, not grey-market evasion.
Western Redlines Create a Global Market Vacuum
The release of LongCat-2.0 arrives at a time when top-tier American laboratories are facing growing regulatory pressure to restrict access to their latest models. Following requests from the U.S. government, OpenAI was forced to limit access to its new GPT-5.6 models, while Anthropic recently restricted access to its latest Claude Fable 5 and Claude Mythos 5 models during ongoing cybersecurity reviews.
These defensive regulatory maneuvers have left a massive operational vacuum in the global developer market. Startups and enterprise developers in Asia, Europe, and Latin America who previously relied on U.S. APIs are increasingly looking for open-source alternatives that are free from geopolitical restrictions. By open-sourcing a highly capable, 1.6-trillion-parameter model with a 1-million-token context window on Hugging Face, Meituan is positioning itself to capture this underserved market. This open-source strategy allows Chinese models to build global developer ecosystems, ensuring their architectures become the foundation for the next generation of AI applications.
The Economics of the Open-Source Pivot
For Meituan, a company primarily known as an on-demand delivery giant rather than a traditional software firm, the decision to open-source LongCat-2.0 represents a calculated business strategy. The domestic AI market in China has turned into a brutal price war, with tech giants like Alibaba, Tencent, and Baidu slashing API costs for their models to near-zero levels. This intense competition has made it difficult for smaller AI startups to survive on API revenues alone.
By open-sourcing the model, Meituan bypasses the direct monetization trap. Instead of trying to sell API access, the company is using LongCat-2.0 to attract a massive community of developers who will build tools, optimizations, and integrations around the model. Meituan can then integrate these innovations back into its core consumer apps, improving its in-house coding efficiency and automating its logistics, delivery dispatch systems, and customer service operations. Furthermore, by proving that it can train and run advanced models entirely on domestic hardware, Meituan has insulated its broader tech operations from any future U.S. sanction reviews, securing its technological future.
Conclusion
Meituan’s open-source release of the 1.6-trillion-parameter LongCat-2.0 model represents a significant shift in the global artificial intelligence landscape. By completing the entire pre-training and inference cycle on a 50,000-chip domestic ASIC cluster, the company has shattered the assumption that frontier-class AI development requires high-end U.S. silicon. Supported by software-level optimizations like the Huawei Collective Communication Library, Chinese tech giants are proving capable of building high-performance supercomputers using alternative, homegrown hardware platforms.
While the United States continues to tighten its technology curbs, these restrictions have inadvertently accelerated the development of a self-sustaining, independent semiconductor and AI ecosystem in China. As Western labs restrict access to their proprietary models, Chinese open-source architectures are stepping into the vacuum, capturing global market share on platforms like Hugging Face. The release of LongCat-2.0 shows that the global race for AI dominance is no longer defined solely by access to a single supplier’s GPUs, but by the ability to innovate across the entire hardware and software stack.
