💰钛媒体•Stalecollected in 12h
Meituan pivots strategy to integrate AI
💡See how a massive local service platform is re-architecting its infrastructure to integrate AI at scale.
⚡ 30-Second TL;DR
What Changed
Stabilizing core revenue streams: food delivery and in-store services
Why It Matters
Meituan's shift highlights how massive local service platforms are leveraging AI to optimize logistics and user engagement to maintain market dominance.
What To Do Next
Analyze Meituan's public technical blog for insights on scaling real-time geospatial AI inference in high-concurrency environments.
Who should care:Developers & AI Engineers
Key Points
- •Stabilizing core revenue streams: food delivery and in-store services
- •Aggressive restructuring of AI infrastructure
- •Navigating a critical period of organizational transition
🧠 Deep Insight
Web-grounded analysis with 18 cited sources.
🔑 Enhanced Key Takeaways
- •Meituan is actively testing LongCat-2.0-Preview, a next-generation large language model with over one trillion parameters, which was trained entirely on domestically developed computing clusters and is considered comparable in capability to GPT-4.
- •The company has open-sourced its 560-billion parameter LongCat-Flash model, built on a Mixture-of-Experts (MoE) architecture, in September 2025, aiming to balance scale and efficiency through dynamic parameter activation and innovations like a 'zero-compute expert mechanism'.
- •Meituan's AI strategy is deeply rooted in 'Physical AI,' focusing on integrating AI into real-world execution capabilities such as logistics, robotics, drones, and autonomous vehicles, with its urban low-altitude drone network entering routine operations in May 2026.
- •Meituan's AI assistant 'Xiaomei' is set to integrate with Tencent's AI assistant 'Yuanbao,' enabling the first 'Agent-to-Agent' communication between two different companies' AI agents in China for local services like food ordering and restaurant booking.
- •The company has made substantial investments in AI infrastructure, including securing billions of dollars in GPU resources since 2024 and re-architecting its systems with Intel Xeon Scalable processors to optimize TensorFlow applications, resulting in a tenfold boost in distributed scalability for recommendation systems.
📊 Competitor Analysis▸ Show
A Markdown table comparing this with competitors (Feature/Pricing/Benchmarks). Return null if not applicable (e.g. op-ed, interview, single-product announcement with no clear competitors).
🛠️ Technical Deep Dive
- LongCat-2.0-Preview: A trillion-parameter large language model (LLM) currently in open testing, trained on domestically developed computing clusters, likely utilizing Huawei's Ascend architecture without NVIDIA hardware.
- LongCat-Flash: An open-sourced 560-billion parameter LLM released in September 2025, employing a Mixture-of-Experts (MoE) architecture. It dynamically activates between 18.6 billion and 31.3 billion parameters per token (averaging ~27 billion) for efficiency.
- MoE Innovations: Features a 'zero-compute expert mechanism' and a shortcut-connected MoE (ScMoE) structure to optimize inference speed and operating costs.
- Training Data: LongCat-Flash was pre-trained on 20 trillion tokens in 30 days.
- AI Infrastructure: Meituan has invested heavily in GPU resources and re-architected its AI infrastructure, partnering with Intel to integrate Intel® Xeon® Scalable processors for TensorFlow application optimization, achieving a tenfold boost in distributed scalability for recommendation systems.
- AI Coding Tools: Developed internal AI programming tools like 'CatPaw,' which reportedly generates approximately 50% of new code, and launched 'NoCode,' an AI-driven platform for non-technical users to create applications through natural language.
- Autonomous Delivery Systems: Deploys autonomous delivery vehicles and drones. By the end of 2024, autonomous vehicles completed 4.91 million orders, and drones completed 450,000 orders. By June 2026, Meituan drones had fulfilled over 900,000 commercial orders globally, with its urban low-altitude drone network in routine operation.
- AI in Operations: Utilizes AI-powered recommendation engines, dynamic pricing algorithms, and machine learning models for its logistics backbone, processing 15 real-time data points per rider, urban traffic patterns, and restaurant preparation times.
🔮 Future ImplicationsAI analysis grounded in cited sources
Meituan will significantly expand its 'Physical AI' capabilities, integrating its advanced LLMs with its autonomous delivery and robotics infrastructure.
The company's strategic focus on 'Physical AI' and its substantial investments in drones, autonomous vehicles, and proprietary LLMs like LongCat position it to create a unique, end-to-end intelligent service ecosystem that bridges the digital and physical worlds.
The 'Agent-to-Agent' integration with Tencent's Yuanbao will establish a new paradigm for cross-platform AI collaboration in China's local services market.
This partnership allows Meituan to leverage another major AI assistant's user base for fulfillment, shifting competition from traffic entry points to the underlying service capabilities and potentially influencing future industry collaborations.
Meituan's open-sourcing of its LongCat-Flash model will accelerate the development of efficient Mixture-of-Experts (MoE) architectures within the broader AI community.
By making its model weights and source code available, Meituan aims to attract talent, build a developer ecosystem, and explore new application scenarios, thereby enhancing its technical influence and potentially setting industry standards for MoE models.
⏳ Timeline
2010-03
Meituan founded as a group-buying website.
2015-10
Merged with Dianping to form Meituan-Dianping, consolidating market leadership.
2023-10
Partnered with Intel to re-architect AI infrastructure using Intel Xeon Scalable processors for TensorFlow optimization.
2024-02
Underwent significant organizational restructuring, integrating 'to-home' and 'in-store' business groups.
2025-05
Launched 'NoCode' AI-driven coding tool and was testing 'CatPaw' for internal developers.
2025-09
Released and open-sourced LongCat-Flash, a 560-billion parameter Mixture-of-Experts (MoE) large language model.
2026-04
Launched open testing for LongCat-2.0-Preview, a trillion-parameter large model trained on domestic computing infrastructure.
2026-06
Announced integration of its AI assistant 'Xiaomei' with Tencent's 'Yuanbao' for local services.
📎 Sources (18)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 钛媒体 ↗