美团扔出万亿参数大模型 本地生活赛道今夜无眠

This article centers on Meituan's official launch of LongCat-2.0, the industry's first trillion-parameter large language model (LLM) with its full training and inference pipeline completed entirely on domestic Chinese computing power. The model is set to bring a brand-new local lifestyle service experience to general consumers. Key takeaways are as follows:

1. Model specifications and performance: LongCat-2.0 has a total of 1.6 trillion parameters, trained on over 30 trillion tokens of pre-training data. It natively supports a 1 million token long context window, runs inference at 100 tokens per second, and costs as low as ¥5 per million tokens. It outperforms leading global models including GPT and Claude on multiple core capability benchmarks, and its total inference volume already ranks among the top three globally.

2. Practical improvements for end users: After integrating LongCat-2.0 into Meituan's AI assistant, the tool will no longer only provide content recommendations. It can directly respond to users' local lifestyle needs and complete end-to-end transactions from inquiry to hotel booking, ticket purchase and food ordering, requiring no manual navigation from users. In the future, Meituan's services will also be accessible directly via WeChat AI agents, enabling an even smoother user experience.

Meituan's new LLM launch reveals shifting consumer and industry trends in the AI era for the local lifestyle sector, offering key insights for brands. Core takeaways are as follows:

1. Trends in consumer experience: AI will restructure the local consumption journey, enabling AI assistants to close the full loop from demand response to transaction directly. Voice input is expected to become the dominant interaction mode going forward. Brands need to adapt their display and transaction logic for AI-native entry points, and start integrating with AI ecosystems early.

2. Guidance for brand and business strategy: Meituan has adopted a pragmatic AI investment strategy that explicitly avoids reckless spending, focusing instead on ROI and business synergy. This approach is suitable for brands building out their own AI capabilities to prevent irrational over-investment. In addition, Meituan's decision to package and open its capabilities as APIs for third-party AI agents points to new channel opportunities for brands, which can tap into traffic by integrating with cross-platform AI ecosystems.

Meituan's AI expansion offers clear opportunity signals and strategic guidance for local lifestyle sellers. Key takeaways are as follows:

1. New traffic and growth opportunities: As one of the first internal testing partners to integrate with WeChat's AI ecosystem, Meituan will soon allow users to access merchant services directly via WeChat AI agents. Meituan has also packaged its capabilities as APIs for agents across all platforms to call on. Sellers can gain more cross-platform exposure and transaction opportunities by leveraging Meituan's AI ecosystem, and should adapt to AI entry point rules ahead of time.

2. Risk warnings and strategic guidance: Meituan founder Wang Xing explicitly notes that AI entry points will not bring disruptive change in the short term, so sellers should not abandon their core businesses to go "all in" on AI blindly. Instead, they should expand their AI footprint in step with platform progress. AI investments should also be kept within a seller's financial capacity, with priority given to projects that boost ROI and improve operational efficiency, to avoid risks from irrational spending.

Meituan's practice of training a trillion-parameter LLM with an end-to-end pipeline on domestic Chinese computing power offers valuable insights for factories advancing digital transformation and AI deployment. Key takeaways are as follows:

1. Supply chain guidance for digital and AI deployment: LongCat-2.0 was fully trained on 50,000 to 60,000 domestic AI accelerator cards, after Meituan overcame core challenges including ten-thousand-card level fault tolerance and improved computing utilization. This achievement confirms that domestic Chinese computing infrastructure can already support large-scale LLM R&D, meaning factories do not need to over-rely on foreign computing power to deploy AI, reducing the risk of supply chain disruptions.

2. Guidance on deployment strategy and business opportunities: Meituan has shifted from early-stage AI experimentation to pragmatic, ROI- and business synergy-focused investment, an approach that is also suitable for factories pursuing digital transformation to avoid wasteful over-investment. Second, as the AI agent ecosystem matures, factories producing lifestyle-related and locally adapted products can access new sales channels and unlock incremental market growth by integrating with AI agent ecosystems. Meituan's collaborative development model with domestic computing chip manufacturers also serves as a reference for partnerships between factories and technology companies.

Meituan's new LLM launch reveals emerging industry trends for AI and local lifestyle services, and outlines solutions to common industry pain points, offering the following key insights for service providers:

1. Industry trend outlook: AI agents will become a critical customer group going forward, and "to-Agent" (To A) services will emerge as a new growth driver. In terms of interaction, voice input will gradually replace part of text input to become the mainstream entry point. At the same time, domestic Chinese computing infrastructure is now mature enough to support large-scale LLM R&D, making this a viable path for development that service providers can prepare for ahead of time.

2. Guidance for addressing customer pain points: The industry has long struggled with high inference costs and insecure computing supply chains. LongCat-2.0 introduced a pioneering zero-calculation expert mechanism that enables token-level dynamic computing allocation, cutting inference costs to ¥5 per million tokens. Its full-stack reliance on domestic Chinese computing also achieves full autonomy and control. These technical approaches can serve as a reference for AI service providers looking to optimize their own products. In addition, Meituan's open API model creates new collaboration opportunities for B2B service providers looking to integrate local lifestyle capabilities.

Meituan's AI deployment practice offers multiple insights for platforms pursuing AI transformation and new business expansion. Key takeaways are as follows:

1. New market demands for platforms: Beyond serving C-end consumers and B-end merchants, serving AI agents has become a new market need. Platforms should package their existing service capabilities into standardized APIs for third-party AI agents to call, unlocking new space for business growth. Meituan has already completed this deployment ahead of peers, integrating with WeChat's AI ecosystem to open up its capabilities, and this model is well worth platform adoption.

2. Guidance for AI strategy and risk mitigation: Meituan makes it clear that AI expansion must have clear financial boundaries, and companies should not engage in irrational spending beyond their capabilities. AI will not deliver disruptive change in the short term, so excessive speculation is unnecessary. Platforms should pursue a steady path of continuous deployment, pragmatic investment and ROI-focused development to avoid excessive financial pressure. On the computing side, prioritizing domestic computing infrastructure and advancing collaborative development between models and chips solves supply chain dependency risks and improves security, and Meituan's experience training large models on ten thousand cards of domestic computing is highly valuable as a reference.

This article discloses the latest developments in AI and the local lifestyle industry, providing first-hand research material for industry analysts. Key takeaways are as follows:

1. New industry developments: China is now home to the first trillion-parameter LLM with a full training and inference pipeline completed entirely on domestic Chinese computing power, validating that domestic infrastructure can support the training of ultra-large models. The industry has currently taken two approaches that both arrive at the same conclusion of domestic computing viability: full-stack migration of foreign architectures, and building an all-domestic process from scratch. Leading companies have shifted from early-stage capital-intensive experimentation to a pragmatic investment phase centered on ROI and business synergy, marking the industry's entry into a more rational new stage of development.

2. New business models and research directions: To-Agent services have emerged as a new business track, where platforms open their service APIs to third-party agents, creating an entirely new business model. AI is also restructuring the full service loop for local lifestyle, shifting from recommendation-focused interaction to an end-to-end transaction execution loop. All of these are new industry directions that merit in-depth research, and Wang Xing's judgments on AI entry formats and AI investment timelines also provide valuable first-hand perspective for research.