In 2026, ordering takeout via AI remains one of the most popular agent experiences for the general public. If there's anything more appealing, it's getting free takeout through AI. Starting yesterday, February 6, Alibaba's Tongyi Qianwen app officially launched its "Spring Festival ¥3 Billion Free Order" campaign. The initiative immediately sparked a surge on Taobao's flash delivery service and at offline bubble tea shops, with the buzz visible across social media, news reports, and daily life.
While some have joked that China's AI landscape ultimately revolves around red envelopes and milk tea, from Yuanbao to Qianwen, the real takeaway is that nearly everyone—from colleagues to friends back home—was trying AI-powered takeout for the first time, often using the phrase, "Qianwen, order me a milk tea." The excitement has continued into today, with users still receiving notifications about high order volumes and delivery constraints.
But where is Meituan, the company behind the Xiaomei app? It has remained largely invisible in this Spring Festival AI showdown. Many may not even be aware that Xiaomei, Meituan's standalone AI agent product, also supports mobile takeout orders, train and flight ticket purchases, and features an AI search function called "Ask Xiaotuan" within the main Meituan app. Despite this, Meituan's AI presence is perceived as faint, and Xiaomei has yet to achieve widespread recognition.
The question is whether Meituan is unable or unwilling to compete—a crucial distinction. On February 6, Meituan released LongCat-Flash-Lite, an open-source model emphasizing agent and coding capabilities. What’s less known is that since its initial launch in September last year, Meituan's self-developed LongCat series of large language models has been continuously upgraded. In fact, the LongCat models have earned respectable evaluations within AI circles, and the user experience on Xiaomei is quite solid.
On September 1, 2025, Meituan released its first self-developed LLM, LongCat-Flash-Chat. Just ten days later, it introduced Xiaomei, an AI-native application designed as a daily life assistant. As a standalone agent product, Xiaomei already possesses comprehensive life-service capabilities. With a single command, users can order food, book restaurants, purchase train tickets, check hotels, and plan trips—all integrated directly into Meituan’s extensive local services network rather than relying on third-party APIs.
In terms of functionality and user experience, Xiaomei is competitive, even outperforming many general-purpose AI agents. Meanwhile, Meituan's main app later incorporated an AI search feature, "Ask Xiaotuan," allowing users to search for restaurants, find deals, and make decisions through natural language.
Over the past six months, Meituan's LongCat series has undergone multiple iterations, with at least nine models released, expanding from text to full multimodal support (image, video), and enhancing reasoning and agent capabilities. In January 2026, the company launched LongCat-Flash-Thinking-2601, an open-source model that topped leaderboards in tool invocation. The newly released LongCat-Flash-Lite is a lightweight Mixture-of-Experts model optimized for developers and agent scenarios, with a focus on agent execution and coding proficiency.
It is worth noting that LongCat-Flash-Lite has a total parameter count of 68.5 billion, with 2.9–4.5 billion activated parameters. It supports context lengths up to 256K via YaRN and integrates an N-gram memory table, outperforming baseline MoE models of equivalent size, particularly in agent and coding tasks. The model is specially optimized for tool use and task decomposition, emphasizing real-world execution—such as multi-step external service calls, complex instructions, and end-to-end workflows. It also excels in code generation and comprehension, enabling more stable programming and automation.
These enhancements point toward a clear goal: enabling large models to power practical, task-oriented agents. This has been the consistent technical direction of the LongCat series. Unlike many models focused on general dialogue and content generation, Meituan prioritizes execution efficiency, cost control, and deep integration with business systems.
Looking back, Meituan’s strategy becomes clearer. The September 2025 release of LongCat marked the company's formal entry into the LLM race. Around the same time, Xiaomei entered public testing, serving as the primary consumer-facing vehicle for LongCat. Over the following six months, Meituan maintained a steady but low-profile update cycle: introducing reasoning-enhanced models, releasing agent task benchmarks, and rolling out image and video generation models to build out multimodal capabilities.
By 2026, this trajectory became even more defined. The LongCat series continued to strengthen reasoning and tool-calling abilities while integrating more deeply with core services like recommendation and merchant systems. In other words, Meituan’s AI development has not stalled—it has been progressing steadily, albeit away from the public spotlight.
This低调 approach creates a perception gap: based on social media alone, one might assume Meituan is absent from both the Spring Festival AI battle and the broader AI competition. Yet, from a model and product perspective, the company has been actively building out its foundation models, agent products, and business integrations over the past half-year.
The internet industry once popularized the term "dimensionality reduction attack." Today, AI may represent a new form of disruption for traditional food delivery. Qianwen’s viral free-order campaign is a case in point. In terms of interaction, traditional apps require users to browse through numerous recommendations, compare options, and spend significant time deciding. AI-driven ordering simplifies this by offering a few tailored suggestions and allowing quick revisions, reducing decision fatigue.
Qianwen’s success is undoubtedly aided by savvy marketing, building on Taobao Flash’s earlier free milk tea promotions. The ¥3 billion campaign isn’t just about subsidies—it directly links AI usage with free orders: to get the discount, users must complete the order through Qianwen. A simple voice command, "Order me a milk tea," handles selection and checkout, combined with free offers and sharing mechanisms that naturally encourage social diffusion.
For many, this is their first time using an AI app to fulfill a daily need. The design encourages adoption not because users "want to try AI," but because they "want milk tea." In contrast, most people remain unaware that Meituan (via Xiaomei) also supports AI takeout, let alone other functions.
This gap stems partly from timing. Over the past year, Meituan has been embroiled in a fierce subsidy war triggered by Taobao Flash, resulting in an adjusted net loss of ¥16 billion in Q3 2025. Earlier, Taobao Flash had acquired Ele.me. With resources diverted to competition and being a relative newcomer to large models, Meituan faced clear capability gaps. Although AI was designated a "core initiative" early in the year, it wasn’t until September that the company delivered tangible outputs.
In this context, AI is treated as a foundational capability requiring sustained investment. That explains Meituan’s intensive iterations over the past six months in both models and AI products. Beyond what’s been mentioned, the company has also launched AI programming tool NoCode and AI IDE CatPaw—based on LongCat—aligning with CEO Wang Xing’s three AI pathways outlined in March 2025: AI at Work, AI in Products, and Building LLM.
Catching up requires resources, talent, and time. How far Meituan’s AI can go remains to be seen, but LongCat’s performance is already impressive, appearing on several international open-source leaderboards and receiving praise on platforms like X (formerly Twitter), YouTube, and Hacker News. In January, LongCat introduced LoZA (LongCat ZigZag Attention), a novel sparse attention mechanism that slashes inference costs for 1M context windows by 90%, hailed by developers as a "VRAM lifesaver."
Still, model success does not guarantee product success. As the LongCat series continues to improve in reasoning and agent capabilities, Meituan is gradually narrowing the model gap. Once underlying models and tool invocation mature further, combined with Meituan’s vast local services network and fulfillment system, the company will be poised to formally join China’s AI battle—allowing more users to discover Xiaomei and related products.
A clear trend in 2026 is the rise of agentic AI. Unlike conversational AI confined to screens, it embeds directly into practical tasks: ordering food, booking tickets, finding hotels, writing reports. Whoever enables more users to hand their daily needs over to AI agents will seize the initiative in the AI era. This represents an opportunity for Alibaba, for Meituan, and for many other internet firms.
Comments