Meituan's Wang Xing: AI Investment Won't Exceed Financial Capacity; "Xiao Tuan" Feature Shows No Explosive Results Yet
Claire Weston
Meituan CEO Wang Xing drew a clear line at the company's AGM: AI investment will stay within cash-flow limits, with no irrational overspending. He also acknowledged the in-app 'Xiao Tuan' AI feature has shown no breakout results so far — signaling a shift from land-grab spending to ROI-first discipline.
What spending line did Wang Xing draw for AI?
Wang Xing stated explicitly: Meituan will invest in AI only within its own cash-flow capacity — no irrational bets beyond financial means.
This means → Meituan's AI strategy has moved past the early "grab a seat first" capital-deployment phase into one driven by capital efficiency and return on investment (ROI).
In plain terms = the company isn't quitting AI — it's insisting every yuan spent must show a visible return. The burn-to-claim-territory era is over.
How is the 'Xiao Tuan' AI feature performing?
Wang Xing was candid: the grey-test 'Xiao Tuan' AI feature inside the main app has "not yet shown any breakout results."
He also predicted that end-user interaction will shift toward voice: "Typing will decrease; voice will increase."
This means → Meituan is cautious about AI's near-term penetration of local-services scenarios. Whether the core business can find a natural entry point through voice interaction is the key validation milestone ahead.
Are the early-stage investments paying off?
In embodied intelligence — AI that gives robots physical form to act in the real world — Meituan-linked funds hold 9.65% of Unitree Robotics, making Meituan its largest outside institutional shareholder. Unitree cleared its STAR Market IPO review on June 1 this year, with a post-offering valuation of roughly RMB 42 billion.
In the large-model space, Meituan holds 3.86% of Zhipu AI. After Zhipu's Hong Kong listing in January, its market cap peaked at HK$650 billion.
This means → with management stressing capital efficiency, the paper gains on these two stakes represent potential non-recurring income — money invested early is starting to flow back.
Why is in-house AI R&D limited to the application layer?
Meituan has explicitly opted out of the arms race in foundational large language models (LLMs), concentrating R&D resources on the application layer and automation toolchains.
In plain terms = rather than building a base model from scratch, Meituan takes models others have built and turns them into user-facing products — cheaper, and closer to monetization.
On June 9, Meituan's Lightyear Beyond team launched Tabbit 1.0, an AI-native browser integrating its in-house LongCat model alongside third-party models including DeepSeek, Zhipu GLM, and Kimi. It is positioned as a cross-app, cross-webpage task-execution gateway.
How is Tabbit performing?
Tabbit's agent task-completion rate — the share of cross-webpage actions the AI successfully executes on behalf of users — has risen from 53.1% during the March beta to 91.8% today.
The product uses tiered pricing: the standard tier is free; the pro tier costs RMB 9.9 per week, offsetting compute costs.
This means → Meituan is trying to make AI products self-sustaining through a "free tier for adoption + paid tier for cost coverage" model, rather than subsidizing them from the core business.
How is the local-services core business plugging into AI?
Meituan is running low-cost AI grey tests across local services, including an errands Skill that connects to third-party AI assistants and a 'Xiao Mei' agent deployed on platforms such as Tencent Yuanbao.
This reflects a deliberate approach: instead of rebuilding its own entry point, Meituan embeds its services inside others' AI ecosystems at minimal cost.
Wang Xing's caution, paired with his voice-interaction prediction, signals that Meituan is waiting for one thing: the day a user saying "order me some takeout" feels more natural than tapping a button — that is when AI truly enters local services.
Content is for reference only, not financial advice.