Summary of Key Points
On July 6th, Meituan open-sourced its new generation large model, LongCat-2.0 (including the model weights, inference engine, and technical documentation). Meanwhile, domestic chip manufacturers such as Huawei Ascend and Moore Threads also completed the adaptation of this model for inference. With 1.6 trillion parameters, LongCat-2.0 is capable of processing millions of characters of text. The core improvements lie in three areas: parameter specifications, domestication of computing power, and scenario-adapted agents. Meituan views AI as a "must-answer" issue, actively investing in it (using 50,000 domestic acceleration cards for training) while maintaining a rational approach (staying within its financial capabilities). Its large model primarily serves as a foundation for internal B2B tools (such as merchant services and research and development efficiency improvements), with limited对外开放 to build barriers in the local lifestyle sector.
Detailed Explanation
1. Open-sourcing + Domestic Chip Adaptation: Breaking the "neck-sticking" issue, making AI more accessible
By open-sourcing LongCat-2.0, Meituan has made the model's "blueprints" and "operating tools" freely available for developers and businesses to use in their applications. More importantly, domestic chip manufacturers have adapted it accordingly. For example, Huawei Ascend's A2 chip with 64GB of video memory per card solved the deployment challenge of a model with 1.6 trillion parameters by using a "PD separation architecture" (separating the model's computing core from data storage, which can be called upon only when needed, thus not consuming all the video memory).
The significance of this is that large models previously relied on foreign chips (such as NVIDIA). Now, using domestic chips not only reduces costs but also eliminates the risk of being constrained by foreign technology. The combination of open-sourcing and domestic adaptation makes AI more affordable and usable for a wider range of companies, fostering the development of the domestic AI ecosystem.
2. LongCat-2.0 Improvements: Larger parameters, better text processing capabilities, and smarter functionality
Compared to previous versions, LongCat-2.0 has made three significant advancements:
- Parameter Specification Upgrade: With 1.6 trillion "sparse parameters" (not all parameters are active simultaneously; only the relevant ones are called upon as needed), the model enhances its ability to handle complex logic (e.g., calculating optimal delivery routes for merchants) and long-duration tasks.
- Completely Domesticated Computing Power: Training with 50,000 domestic acceleration cards eliminates the need for foreign chips, reducing costs and enhancing supply chain autonomy.
- Scenario-Adapted Agents: The model can process over 1 million characters of text (equivalent to reading a novel) and has optimized code processing and automated workflows (e.g., automatically generating marketing plans for merchants). Its agent capabilities are nearly on par with the international top model, Claude Opus (with a score of only 0.2 points behind).
3. Meituan's Approach to AI: A strategic investment without reckless spending
Shareholders were concerned about Meituan's aggressive AI investments, but Wang Xing clarified that AI is a "must-answer" issue for all companies. Meituan already uses AI in its products and operations, such as the "Xiao Tuan AI Assistant" that helps users search for information, and agent tools that optimize internal processes. The CFO mentioned the company has 65 billion yuan in investment reserves (e.g., holding shares in companies like Ideal and Zhipu), ensuring investments stay within financial limits to keep up with trends without taking unnecessary risks.
4. The Value of LongCat-2.0 for Meituan: Internal cost savings and external barriers
LongCat-2.0 serves as a technical foundation for Meituan's internal B2B services:
- Internal Efficiency: It supports tools like "Kangaroo Advisor" (helping merchants analyze business data) and "Intelligent Store Manager" (automatically responding to customers), and AI-assisted coding tools speed up development, reducing operational costs.
- External Barriers: By using domestic computing power, inference costs are lower, attracting developers to create local lifestyle applications. Combined with Meituan's vast amount of data (on transactions, deliveries, and merchant operations), it provides small and medium-sized merchants with precise location recommendations and marketing advice—a unique advantage in the local lifestyle sector.
5. Industry Impact: Promoting collaboration in the domestic AI ecosystem
Meituan's open-sourcing of the large model encourages more companies and developers to innovate based on it, while domestic chip manufacturers find new use cases for their hardware. This combination establishes a domestic chain for models and chips, reducing reliance on foreign technology and facilitating the integration of AI into real economy sectors like local life and retail, making it more accessible to everyday people.
In summary, Meituan's move is not only an important step in its own AI strategy but also a boost for the domestic AI ecosystem. By lowering barriers through open-sourcing and using domestic solutions, it turns AI from a high-tech concept into a practical tool that can truly benefit merchants and users.