Speeding Up AI Development: Meituan Seeks a 'Springboard' for Its Food Delivery Empire

Meituan Charges Full Throttle Toward the Large Model Summit

Recently, 36Kr reported that Pan Xin, a former researcher at Google DeepMind and head of ByteDance's visual large model AI platform, has recently joined Meituan as the leader of its multimodal AI innovation business.

In addition to attracting top technical talent, Meituan is also expanding its team of AI researchers and engineers. According to AIbase, throughout 2025, Meituan has been consistently recruiting AI training experts, targeting teams from Alibaba's Tongyi and Seed, as well as Tencent's Hunyuan. The interview criteria prioritize candidates' abilities in "real-world scenario data loops" and minute-level decision-making.

Meanwhile, a source close to Meituan revealed to Super Focus that since fully embracing domestic technology routes in 2023, Meituan has acquired a substantial number of Ascend computing cards for training its internal large models. In the second quarter of this year, the company prioritized "basic model training" as its top project. Wang Puzhong even disclosed during an event that Meituan invests over 10 billion yuan annually in AI, underscoring the company's strategic emphasis on and commitment to the AI sector.

So, how capable are Meituan's large models under such heavy investment? To what extent have they enhanced Meituan's business? And why is Meituan so persistent in its pursuit of AI?

01

Annual Investment Exceeds 10 Billion: Meituan Embraces AI Like Tech Titans

The pinyin for "takeout" (food delivery) naturally contains two "ai"s, seemingly foreshadowing that Meituan, a super local life service provider, is destined to be deeply intertwined with artificial intelligence. And indeed, after quiet investment and continuous iteration, Meituan is emerging as a formidable contender in China's large model arena.

In September, Meituan officially released and open-sourced its first large language model, LongCat-Flash-Chat, along with the LongCat-Flash-Thinking model, which focuses on reasoning capabilities. The launch of these large language models laid the groundwork for subsequent applications in complex business scenarios.

Next, Meituan shifted its focus to the multimedia domain, releasing the LongCat-Video video generation model in October, marking its entry into high-dimensional multimedia synthesis and understanding. This move explores how AI technology can enhance efficiency and innovation in short video and marketing content creation.

In November, Meituan's technological stack advanced toward a unified cognitive architecture with the release of the full-modal model LongCat-Flash-Omni. This model aims to seamlessly integrate and understand multi-source heterogeneous data, such as language and vision, propelling Meituan toward a general intelligence agent capable of complex semantic reasoning and multi-channel decision-making.

As the year drew to a close, Meituan concentrated on strengthening its visual foundation capabilities, open-sourcing three image models, including LongCat-Image, in December. This matrix will significantly empower AI applications in Meituan's core businesses, such as merchant image processing, product display optimization, and user-generated content.

This means that within a single quarter, Meituan not only rapidly completed the construction of a full-modal AI foundation spanning text understanding, vision, and video but also achieved this while fully embracing domestic technology routes at the underlying computing power level. This rapid and robust self-developed capability highlights Meituan's core technological prowess in the AI direction.

However, such comprehensive AI capabilities are not built on thin air but are directly empowering the local life service ecosystem through specific AI agents and SaaS tools.

In terms of enhancing the C-end user experience, the "Xiaomei" App, Meituan's first AI agent product, marks the company's attempt to revolutionize traditional selection modes with conversational AI.

Powered by self-developed models, it enables more natural and fluid multimodal interactions, providing users with a one-stop, efficient, and intelligent local life service experience, creating a small dining ecosystem within Meituan.

On the merchant efficiency front, Meituan has translated its large model capabilities into business tools.

The free availability of SaaS tools like "Daishu Staff Officer" (Kangaroo Advisor) and "Intelligent Shopkeeper" leverages Meituan's vast data and LongCat large model capabilities to deliver "data intelligence" for refined operations of small and medium-sized merchants. This not only helps merchants improve management efficiency and profitability but also solidifies Meituan's irreplaceable position in the local merchant ecosystem through technological empowerment.

It can be said that Meituan's AI investment is not merely theoretical; its results have deeply penetrated and empowered every aspect of the local life service ecosystem through specific agents and SaaS tools.

However, despite AI being deployed in business applications and possessing long-term strategic value, the heavy investment exceeding 10 billion yuan annually has led to significant market criticism, especially amid recent major business turmoil caused by intense competition.

Currently, Meituan is facing fierce competition with e-commerce giants in the instant retail sector. At a time when substantial funds are needed for market subsidies, rider capacity, and merchant incentives to compete for users and market share, the "billion-dollar" investment in large model research stands out, naturally triggering internal anxiety about resource allocation efficiency.

Meanwhile, there are also deep-seated doubts about the return on investment within the company. Some employees worry that certain large model applications are "AI for the sake of AI" rather than genuinely addressing business pain points. For example, will C-end users, in a highly scenario-specific App, truly abandon familiar selection interfaces for early-stage conversational AI agents? And if such habits are cultivated, why not use general agents from other hardware manufacturers instead?

Furthermore, self-developed large models entail significant computing power investment, continuous hardware upgrades, and maintenance costs; top AI scientists command high salaries, and talent competition is fierce; adopting a domestic card path requires higher technical standards, and adaptation is highly dependent on computing card manufacturers...

Such massive investments, technical barriers, and uncertainties in business models collectively contribute to the market's cautious attitude toward Meituan's AI strategy.

02

Under 10 Billion AI Investment: Meituan's Relentless Pursuit of the Top Tier

Behind the seemingly aggressive investment lies Meituan's unwavering determination not to settle for a "second-tier" position.

In the internet rankings, Meituan is often habitually categorized as a vertical giant, but internally, it consistently evaluates itself against "first-tier" standards in both strategic vision and technological breakthroughs.

This self-positioning stems not only from confidence in surviving the "food delivery wars" but also from a clear understanding of the survival rules in the internet's second half: pure scale expansion has peaked, and only by mastering the most cutting-edge foundational technology can one maintain an absolute leading position amidst giant competition.

For Wang Xing and Meituan, a billion-dollar AI investment is not an expensive embellishment but a ticket to the competitive arena of the next decade. It is not just a tactical defense but a strategic leap forward.

This obsession with a "first-tier" identity is essentially an extreme thirst for growth certainty.

Looking at Meituan's core fundamentals, while its food delivery and flash sale businesses have established high moats, it is undeniable that the penetration rate of local life services is gradually approaching its theoretical limit.

Even without the "Battle of the Gods" in instant retail in 2025, Meituan has effectively completed its deep harvest of the industry. As the market enters a mature phase, even if Meituan can maintain a dynamic balance through extreme efficiency in fierce attrition wars, it cannot hide the fact that its growth space is narrowing.

The traditional "manpower + ground promotion + algorithmic scheduling" model has hit a ceiling. Without a dimensional leap, Meituan will inevitably fall into a muddle of low-level repetitive competition. In this context, finding a second growth curve is no longer an "icing on the cake" option but a life-or-death question.

On Meituan's strategic chessboard, besides its accelerating overseas expansion, technological transformation, especially general artificial intelligence technology centered on large models, has become its heaviest bet on the future.

Meituan knows that overseas expansion is spatial expansion, while technology is temporal reshaping. This obsession with technology is actually an evolution forced by business demands. When food delivery riders' route optimizations are precise to the second and merchant backend auto-dispatching achieves millisecond responses, the dividends of mere digital transformation have been exhausted.

Meituan must fundamentally reconstruct the relationships between people, goods, and scenes through AI large models, upgrading from "simple matching" to "deep decision-making" and even expanding to other scenarios.

This high-intensity investment in technology reflects Meituan's ambition to break free from business constraints and establish new growth paradigms. Even if the market remains skeptical about current return on investment, for Meituan, such strategic "waste" is the necessary cost to maintain its first-tier combat effectiveness.

After all, in the torrent of technological evolution, the greatest risk is never over-investing but finding oneself without a winning hand when a new era begins.

However, lofty ideals often face stark realities: as a giant that grew from "labor-intensive" scenarios like food delivery and in-store services, Meituan's attempt to compete head-on with traditional internet giants possessing deep cloud computing foundations and native technological genes on the large model frontier is no less challenging than changing lanes in turbulent waters.

Moreover, in this costly and time-consuming arms race, the imagination of new businesses essentially depends on the "blood-transfusion" capabilities of old businesses. The large model game may seem like a technological competition, but behind it lies a comprehensive contest of cash flow and scenario landing efficiency.

For Meituan, this means not only being able to invest but also to execute effectively.

Currently, Meituan is in a delicate position of being attacked from both sides: on one hand, its old businesses are engaged in close combat with strong rivals like Taobao Flash Sales in the instant retail sector, with every point of profit accompanied by high marketing and logistics costs; on the other hand, AI infrastructure construction requires billion-dollar sustained investment. This dual-line pressure is an extreme test of Meituan's financial resilience and strategic determination.

But as Wang Xing once advocated in the "infinite game," business competition has no final outcome.

Meituan's choice to deeply cultivate AI is an attempt to leap from low-dimensional traffic competition to high-dimensional technological dominance. Although the current return on investment remains controversial, and there is still a long road ahead in domestic adaptation and model tuning, just like during the food delivery wars, the most difficult choices often yield the most valuable outcomes.

In this era of great change, the only certainty is uncertainty itself, and Meituan has already grasped that ticket to the future. Whether it can ultimately reach the shore, time will provide the most impartial answer.

- END-