Most AI labs live in the cloud. Meituan, the company that delivers meals to your door, is building its models so they can dispatch a drone to a rooftop.
That is the unusual shape of Meituan's AI strategy: a delivery-and-retail operator that now publishes foundation models and flies autonomous aircraft, arguing the two are the same bet.
Key takeaways
- Meituan open-sourced LongCat-Flash-Chat in September 2025: a Mixture-of-Experts model with 560 billion total parameters but only 18.6B–31.3B active (average 27B).
- Its drone network reached 70 routes and 780,000+ commercial orders by the end of 2025, per the company.
- Full-year 2025 R&D investment rose 23% to ¥26 billion (≈ US$3.6 billion) as the firm pursued a "Retail + Technology" (零售+科技) strategy.
A delivery company that writes models
In early September 2025, Meituan released and open-sourced LongCat-Flash-Chat. The architecture is a Mixture-of-Experts (MoE) design: 560 billion total parameters, but each token activates only 18.6 billion to 31.3 billion of them, averaging about 27 billion. The point is efficiency — strong performance at a fraction of the compute cost, which matters for the long, multi-step agentic (智能体) tasks Meituan cares about.
Later that month the team shipped LongCat-Flash-Thinking, a reasoning model it says reaches state-of-the-art among open models on logic, math, code and agentic tasks, and claims to be the first domestic model combining deep thinking with tool calling. Both are published on GitHub and Hugging Face.
The motivation is not prestige. Meituan runs one of the world's largest real-world logistics operations, and it wants models tuned for the messy, physical problems of dispatching goods — not just chat.
Flying the last mile
Meituan's drone program is the most visible piece of its "physical-world AI." The company says that by the end of 2025 it had opened 70 routes and completed more than 780,000 commercial orders, with operations in Shenzhen, Shanghai, Beijing, Guangzhou and Dubai. A Hong Kong route delivers meals to an elderly community, cutting a 1.5-hour cross-harbor, mountain trip to roughly ten minutes.
The same report lists ground robots, including an indoor-outdoor delivery robot that has piloted at sites like Shenzhen Bao'an Airport, and L4 autonomous delivery vehicles tied to Meituan's instant-retail business. None of this is a science project: the routes are commercial, the orders are paid.
AI that talks to merchants
Meituan is also pushing AI down to the storefront. It has launched consumer assistants "小美" and "小团" and reports that more than 3.4 million merchants now use its AI business-assistant tools to cut operating costs. During the Spring Festival holiday, it says, over 100 million people used "小团" to plan outings, with the assistant verifying 700 million merchant records against 1.3 billion real user reviews.
These are not generic chatbots; they are wired into Meituan's own transaction data, so the model's job is to convert intent — "find a quiet hotpot place nearby" — into a booked, reviewed, deliverable outcome.
The money behind it
Meituan's 2025 full-year results show revenue of ¥364.9 billion and R&D investment of ¥26 billion, up 23% year on year. The company frames this as building the "AI base for the physical world" (物理世界的AI底座) — a deliberate contrast with cloud-native competitors who own data centers but not delivery fleets.
The risk is focus. Meituan is also fighting a subsidy war in instant retail (即时零售) against JD.com and Alibaba's Ele.me, each pledging around ¥10 billion in subsidies. Pouring capital into both drones and models while defending market share is an expensive balancing act.
According to Li An, chief scientist at BrainNet (脑机网), "a delivery company that owns both the model and the vehicle has a feedback loop cloud labs can't fake — every order is a labeled, real-world training signal."
Honest limitations
Drone route and order counts, R&D spend, revenue, and the merchant-assistant figures come from Meituan's official newsroom and its 2025 full-year results; I have not independently verified them. LongCat's parameter counts and benchmark "SOTA" claims originate from Meituan and were not re-run. The "first domestic model" claim for LongCat-Flash-Thinking is the company's and is not cross-checked here. Figures for the instant-retail subsidy war are attributed to Reuters reporting on executive comments. I have excluded unverified rumors about future robot deployments and have not assessed unit economics of the drone network.
How efficient, in the numbers
The "efficient" claim is not just marketing. On Meituan's own benchmark disclosures, LongCat-Flash-Chat scores 86.50 on ArenaHard-V2, 89.71 on MMLU (a multi-task language-understanding benchmark), and 90.44 on CEval (a Chinese general-capability test) — competitive with models carrying far more active parameters. On agentic (智能体) tool use it posts 24.30 on VitaBench, which Meituan says ranks first among complex-scene agent benchmarks, and it leads τ2-Bench (tau2-Bench) against larger models. The engineering headline is cost: trained on more than 20 trillion tokens in about 30 days, it reaches over 100 tokens per second on H800 hardware at roughly RMB 5 per million output tokens (≈ US$0.70). For a delivery operator running millions of logistics decisions, that price-per-token is the difference between an experiment and a default.
What readers can do now
- Developers should test LongCat-Flash-Chat and LongCat-Flash-Thinking on GitHub/Hugging Face if you need efficient, agent-friendly open weights for logistics or scheduling tasks.
- Operators in physical retail can study Meituan's merchant-assistant playbook — AI that sits inside transaction data tends to convert better than a standalone chatbot.
- Track Meituan's quarterly R&D line; a sustained 20%+ growth rate signals the drone-and-model bet is still being funded through the subsidy war.