Technical Deep Dive
The integration hinges on two key technical layers: WeChat's newly exposed AI Agent interface and Meituan's LongCat-2.0-Preview model.
WeChat's AI Agent Interface: WeChat has released a set of APIs that allow AI Agents to directly invoke mini-program functions—a move that effectively turns the entire WeChat mini-program ecosystem into a callable service graph. The interface is designed around a 'function calling' paradigm, where an agent can send structured requests (e.g., `order_food(location, cuisine, budget)`) and receive real-time responses with pricing, availability, and confirmation. This is architecturally similar to OpenAI's function calling but tailored for the mini-program environment, which has over 8 million registered mini-programs as of early 2025. The latency requirement is stringent: the entire round-trip from agent intent to service confirmation must complete under 2 seconds for a seamless user experience.
Meituan's LongCat-2.0-Preview: This is a dense MoE (Mixture of Experts) model with a total parameter count exceeding 1 trillion (1.02T estimated). It was trained entirely on domestic compute clusters using Huawei Ascend 910B chips, a significant milestone for China's AI sovereignty. The model's architecture uses 64 experts with a top-2 routing mechanism, meaning only ~20B parameters are activated per forward pass, keeping inference costs manageable. LongCat-2.0 achieves a reported MMLU score of 89.1, placing it slightly above GPT-4o (88.7) but below Claude 3.5 Sonnet (88.3) on the same benchmark—though these numbers are from Meituan's internal evaluations and should be treated with caution.
Agent Architecture: The two flagship agents, 'Xiao Tuan' (focused on food delivery) and 'Xiao Mei' (focused on lifestyle services like movies, travel, and appointments), share a common reasoning backbone but have specialized tool-use modules. The system uses a 'plan-then-execute' pipeline: first, the agent decomposes a user's natural language request into a sequence of sub-tasks (e.g., 'order dinner from a Sichuan restaurant near the cinema at 7 PM'), then it queries WeChat's service registry to find the appropriate mini-programs, calls them in parallel where possible, and presents a unified confirmation to the user. This is implemented using a variant of the ReAct (Reasoning + Acting) pattern, with a custom 'service graph' memory that caches frequently used mini-program endpoints to reduce latency.
Data Takeaway: The latency requirement of <2 seconds for agent-to-service calls is a major engineering challenge. Meituan reports that their current system achieves a median latency of 1.4 seconds for single-service calls and 2.8 seconds for multi-service orchestration (e.g., dinner + movie + taxi). The multi-service latency is still above the 2-second threshold, indicating room for optimization.
Key Players & Case Studies
Meituan (Meituan-Dianping) is the clear leader in China's local services market, with over 700 million annual active users and a network of 9 million merchants. Its CEO Wang Xing has been vocal about AI being the 'biggest opportunity since mobile internet.' The company has invested over $3 billion in AI R&D since 2023, with LongCat being the crown jewel. The 'To A' strategy is a direct bet that AI Agents will become a new distribution channel, potentially replacing the need for users to open the Meituan app at all.
Tencent (WeChat) is the platform owner. WeChat's AI ecosystem opening is part of a broader 'WeChat AI' initiative that includes the 'Yuanbao' assistant and various generative AI features. By allowing external agents, Tencent is effectively creating a 'meta-app store' for AI services, where agents compete to serve users. This is a defensive move against ByteDance's Douyin, which has been aggressively expanding into local services with its own AI capabilities.
Competing Approaches:
| Company | Model | Parameters | Local Services Integration | Agent Strategy |
|---|---|---|---|---|
| Meituan | LongCat-2.0-Preview | 1.02T (MoE) | Deep (native) | To A (Agent-first) |
| Alibaba (Ele.me) | Tongyi Qianwen 2.5 | ~500B (dense) | Partial (via Alipay) | To C (Assistant) |
| ByteDance (Douyin) | Doubao 1.5 | ~400B (dense) | Emerging (in-app) | To C (Content-driven) |
| Baidu (Waimai) | ERNIE 4.5 | ~300B (dense) | Weak (partnered) | To B (Enterprise) |
Data Takeaway: Meituan's trillion-parameter model gives it a significant advantage in complex reasoning tasks, but the inference cost is high—estimated at $0.15 per 1M tokens, compared to $0.05 for Alibaba's Tongyi Qianwen. The trade-off between capability and cost will be critical as agent usage scales.
Case Study: The 'Date Night' Scenario
A user tells Xiao Mei: 'I want to take my partner out for a nice dinner near the IFC mall in Shanghai, then watch a movie, and get a taxi home afterward.' The agent must:
1. Query restaurant availability (cuisine, budget, ratings) via Meituan's mini-program
2. Check movie showtimes and seat availability via Maoyan (a Meituan affiliate)
3. Book a taxi via Didi (integrated via WeChat's mini-program)
4. Optimize the schedule to avoid conflicts (e.g., dinner must end 30 min before movie starts)
5. Present a unified itinerary with total cost and one-click confirmation
This requires multi-step reasoning, real-time data fetching, and constraint satisfaction—a task that traditional search interfaces cannot handle. Early testing shows a 73% user satisfaction rate for fully automated plans, compared to 58% for manual planning.
Industry Impact & Market Dynamics
The shift to AI Agent-mediated services has profound implications for the local services market, which was valued at approximately $1.2 trillion in China in 2025.
Distribution Power Shift: Currently, local services are dominated by app-based search and recommendation. AI Agents introduce a new 'zero-click' paradigm where the agent does all the work. This threatens the traditional ad-based revenue model, as merchants may no longer need to bid for keywords if an agent selects based on objective criteria (price, rating, distance). Meituan's response is to create a 'merchant agent API' that allows businesses to bid for agent placement—a new form of 'agent-side advertising.'
Market Share Dynamics:
| Platform | 2024 Market Share (Local Services) | 2025 Est. Market Share | AI Agent Readiness Score |
|---|---|---|---|
| Meituan | 65% | 62% | 9.5/10 |
| Alibaba (Ele.me) | 20% | 18% | 6/10 |
| ByteDance (Douyin) | 10% | 15% | 8/10 |
| Others | 5% | 5% | 4/10 |
Data Takeaway: Meituan's market share is expected to decline slightly as ByteDance gains ground, but its AI agent readiness gives it a strong defensive position. The key variable is whether users trust agents enough to cede control over purchasing decisions.
Funding and Investment: The AI agent space for local services has attracted significant venture capital. In Q1 2025 alone, over $4.2 billion was invested in AI agent startups globally, with China accounting for $1.8 billion. Meituan's internal AI unit has been valued at $12 billion in secondary market transactions.
Risks, Limitations & Open Questions
1. User Trust and Agency: The biggest risk is user backlash against 'agent-driven' decisions. If an agent orders the wrong food or books a bad movie, who is responsible? Meituan has implemented a 'human-in-the-loop' override for all transactions above $20, but this adds friction. The open question is whether users will accept the trade-off between convenience and control.
2. Model Hallucination in Service Context: LongCat-2.0, like all LLMs, can hallucinate. In a service context, hallucination means suggesting a restaurant that is closed, or a movie that is sold out. Meituan reports a 2.3% hallucination rate in service recommendations, which they aim to reduce to <0.5% through real-time verification against the mini-program database. However, verification adds latency.
3. Competitive Response from Alibaba and ByteDance: Alibaba is reportedly developing a 'Tongyi Agent' specifically for Ele.me, but it lags behind in integration depth. ByteDance's Douyin has the advantage of a massive user base and strong AI capabilities, but its local services infrastructure is less mature. The risk is that Meituan's first-mover advantage may be eroded if competitors offer better agent experiences.
4. Regulatory Scrutiny: China's Cyberspace Administration has yet to issue specific guidelines for AI Agents that execute financial transactions. The current regulatory framework for AI (e.g., the 2023 Generative AI Measures) focuses on content generation, not autonomous action. There is a risk that regulators could impose strict liability rules on agents, requiring explicit user consent for every transaction—which would defeat the purpose of automation.
5. Economic Viability: Running a trillion-parameter model for millions of daily agent interactions is expensive. Meituan estimates that each complex multi-service request costs approximately $0.03 in compute. At scale, this could amount to millions of dollars per day. The company plans to offset this through a combination of merchant commissions (higher for agent-sourced orders) and premium subscriptions for 'agent-enhanced' service tiers.
AINews Verdict & Predictions
Verdict: The WeChat-Meituan AI agent integration is the most significant development in Chinese local services since the launch of mini-programs in 2017. It represents a genuine paradigm shift from 'human searches for services' to 'agents negotiate services on behalf of humans.' Meituan's trillion-parameter LongCat-2.0 gives it a technical edge, but the real moat is the depth of its service graph—the ability to orchestrate across 9 million merchants in real-time.
Predictions:
1. By Q3 2026, AI Agents will account for 15% of Meituan's total order volume. This will be driven by the 'Date Night' and 'Quick Lunch' use cases, where convenience outweighs the need for manual selection.
2. WeChat will open its agent ecosystem to third-party developers within 12 months. This will create a 'agent app store' where developers can build specialized agents (e.g., a 'Pet Care Agent' that books vet appointments and orders pet food). This will be the next major platform play.
3. Alibaba will acquire or build a competing trillion-parameter model for local services within 18 months. The current gap in model capability is too large to ignore. Expect a bidding war for AI talent specializing in service-oriented MoE architectures.
4. Regulatory intervention is likely by 2027. The Chinese government will mandate that all AI Agents executing financial transactions must provide a 'human override' option and maintain a transparent audit trail. This will increase compliance costs but also legitimize the industry.
5. The 'To A' strategy will be adopted by other verticals. Expect to see similar agent-first approaches in travel (Ctrip), ride-hailing (Didi), and even healthcare (WeDoctor) within the next two years. The agent becomes the new universal interface.
What to Watch Next: The key metric to track is 'agent stickiness'—the percentage of users who return to the agent for repeat transactions. If Meituan can achieve >60% retention within three months, the paradigm shift is real. If retention is below 30%, the agent may be a novelty rather than a new distribution channel.