Meituan releases LongCat-2.0, a trillion-parameter large model trained on domestic computing power, stemming from complex scenarios like food delivery dispatch, emphasizing data assets and scenario adaptation.
Meituan Quietly Trains a Trillion-Parameter AI Model. Why Is a Food Delivery Company Building Cutting-Edge AI?
Last week, something a bit incongruous happened. Meituan — the platform that delivers your fried chicken and books your massages — quietly released a large language model called LongCat-2.0. It has 1.6 trillion parameters. Not 1.6 billion — 1.6 trillion, where "T" stands for trillion. Globally, you can count on two hands the number of models running at this scale.
Even more unusual: this model wasn't trained on Nvidia GPUs. It completed its full training and inference pipeline on 50,000 domestic AI accelerator cards — the first trillion-parameter model in the world to do so.
Honestly, when I first saw the news, my reaction was: Why is Meituan doing this? Do delivery riders need a trillion-parameter model on their routes? But this isn't that simple.
Food Delivery Is a Hundred Times More Complex Than You Think
Let's start with the "simple" scenario of delivery dispatch. During the lunch rush, within a three-kilometer radius in Beijing, you have hundreds of riders and thousands of orders, all shifting every minute. One restaurant falls behind on orders, a traffic accident just happened on one street, an elevator in an apartment complex broke, a user just changed their delivery address. These variables all happen at once, and the system needs to re-plan every rider's route, estimate delivery time for every order, and assign incoming orders appropriately — all at millisecond speed. In industry terms, this is called "real-time dynamic dispatch," widely recognized as an extremely complex computational problem.
Meituan processes tens of millions of orders every day. This system has run for over a decade, accumulating a vast trove of real-world data: traffic conditions, weather, user tolerance for waiting, restaurant preparation patterns, rider behavior habits... These data share a common trait: they're messy, noisy, and real. You won't find them in textbooks, and they're not in academic datasets. They can only be accumulated through day-in, day-out operation in actual delivery scenarios.
And large models, put simply, grow by being fed data. Whoever's data is closer to the real world ends up with a model that better understands the real world. Meituan isn't building large models to chase a trend — it's because they hold something no one else has.
A Trillion-Parameter Model Fed by Dispatch Data
There's a detail in the LongCat-2.0 release that most people missed. The model's architecture was designed to be more efficient at "Agentic Coding" tasks — enabling AI to work like a real programmer: understanding problems on its own, writing code, executing, debugging, and iterating. Why did Meituan design it this way? Because core systems like delivery dispatch and route planning sit on top of extremely complex engineering codebases. When the business needs to iterate quickly and validate new delivery strategies, an AI that can write its own code is the most direct productivity tool available.
The 1M-token context window follows the same logic. Meituan's codebase runs to millions of lines; a single user's historical behavior records can stretch very long; supply chain history can go back years. For the model to truly understand all of this, it needs to be able to read long content "in one go," rather than "losing memory" every few thousand tokens.
This is why people say internet companies' AI "grows out of their scenarios." Behind every design decision in the model, there's a specific business pain point pushing it forward. This isn't unique to Meituan — ByteDance, Alibaba, and Tencent all take the same approach: feed their own scenario data in, and grow a model suited to their business. ByteDance has Douyin's recommendation scenarios and Volcano Engine, Alibaba has Taobao's e-commerce data and Tongyi Qianwen, and Tencent holds a strong hand too — WeChat's social graph and advertising system capture the daily behavior of a billion users. Every company's path into AI is dug from its deepest moat.
And here's something I think many people haven't fully thought through: does delivery dispatch really need a trillion parameters? Wouldn't a smaller, vertical model be more cost-effective? That's a fair question, but the answer isn't so straightforward. A smaller model is sufficient for high-frequency mainstream scenarios. On a clear weekday on Beijing's Third Ring Road, order volume is stable and traffic is predictable — a moderately sized model can handle it fine. But the real challenge in delivery dispatch is the long tail of scenarios, which is wildly extensive. Heavy rain plus a holiday plus a subway failure plus some viral restaurant suddenly trending on social media — how many times have you seen that combination in a training set? A restaurant shutting down temporarily, multiple riders involved in incidents at once, a neighborhood closing a road for construction — each of these is low-frequency on its own, but when variables combine in pairs, triples, and beyond, the possibilities approach infinity. A smaller model essentially breaks on these combinations: it either produces a nonsensical dispatch plan or simply goes silent. The point of a trillion parameters is to use a sufficiently large parameter space to "remember" those extreme cases. It's not that every parameter is actively working — it's that when an anomalous scenario appears, the model has enough capacity to recognize it and fall back gracefully instead of giving up entirely. That cost is considerable, but Meituan clearly thinks it's worth it.
Why Train Your Own Model When You Could Just Call Someone Else's?
You might ask: with ChatGPT, Claude, and various domestic model APIs available, why not just call them directly? Why spend the enormous money to build a 50,000-card cluster and train your own model? There's more than one reason.
The most direct reason is that the data cannot leave the building. Dispatch system data, consumer behavior, merchant operating data — these are Meituan's most critical business assets, and they involve significant user privacy. Feeding that data into someone else's model is something no serious company would do. To truly teach a model this data, you have to train it yourself.
Then there's latency. Delivery dispatch requires millisecond-level decisions, ad bidding is real-time, and recommendation systems recompute on every refresh. Calling an external model — with several hundred milliseconds per request — simply can't work in these scenarios.
There's also control. With your own model, you can fine-tune it precisely for your scenarios; when an external model updates, your business might break. For core systems, that kind of uncertainty is unacceptable.
But what really convinces me the money had to be spent is this: model capability ends up defining the boundaries of the business. Meituan's EvoCUA (an AI agent that can write its own code) is already in use. The search and recommendation ASX team has built an agent-based technology stack on large models, and the fulfillment AI team is using the same foundation to build a self-evolving operations system. None of this would be possible without having sufficiently strong model capabilities of your own. With a foundation model like LongCat, Meituan isn't just improving delivery time estimates — it's enabling the entire operations system to identify problems and adjust itself. That gap can't be bought.
The 50,000 Domestic AI Accelerator Cards Tell Another Story
Before getting into this, I want to be clear about something, because I think a lot of tech coverage has been skewed toward the good news. A 50,000-card cluster is not an easy decision. Electricity costs, cooling costs, and maintenance costs add up to astronomical numbers. Even more expensive is the talent: engineers who can manage training at this scale are scarce globally, and hiring them is only half the battle — retaining them is another story entirely. This path carries heavy stakes and a long cycle, and a failure at any point could waste months of training. I think moving forward requires not just technical judgment, but a rare kind of organizational resolve. I'm not being a downer — I just think that understanding these costs is necessary to appreciate why what comes next matters.
Back to the detail in LongCat-2.0 that most people skimmed past: 50,000 domestic accelerator cards. This shook the technical community even more than the model itself. Training a trillion-parameter model places extremely demanding requirements on compute. If a single card computes incorrectly, the entire training run can collapse. With 50,000 cards working together, the probability of errors rises exponentially, and you're dealing with a host of engineering challenges around communication bandwidth and scheduling stability. Until now, the industry generally believed domestic accelerators weren't ready for tasks at this scale — the unspoken assumption was that this level of training could only run on Nvidia's A100 or H100. LongCat-2.0 changed that assumption.
Huawei Ascend, Moore Threads, and MetaX all completed inference adaptation on the same day the model was released. Behind this, the software ecosystem is catching up. This isn't just a Meituan story. When leading Chinese internet companies start running their flagship models on domestic accelerators, the entire chain comes alive: compute vendors finally have a credible flagship customer, the toolchain gets stress-tested under real-world, large-scale conditions for the first time, and engineers have a reference implementation to learn from.
Meituan's decision to open-source both the training code and the inference engine at this point helps not only research teams wanting to reproduce the model, but also engineers trying to understand "what domestic accelerators can actually do." And by the way, the open-source move isn't without its own strategic calculus — influence over technical standards and the ability to attract a strong talent ecosystem are often built exactly this way.
Final Thoughts
While writing this piece, I kept circling back to a question: are China's major tech companies going all-in on AI because they genuinely have a vision, or because they're afraid of being called "behind the times" by their peers? After looking at Meituan's case, my initial skepticism shifted.
Meituan's path to large models has plenty of earlier signs if you trace it back. Delivery dispatch optimization, supply chain forecasting, grocery inventory management — these businesses have been algorithm-driven for a long time. For them, large models aren't so much a new track as the natural turn at the next intersection on a road they were already on. The same logic holds for ByteDance, Alibaba, and Tencent. They aren't chasing a goal defined by someone else — each is digging deeper into its own business core.
Of course, to be honest, that doesn't mean they'll all succeed, and it doesn't mean every investment will yield a clear return. Anyone who knows the space understands how hard AI deployment really is. Whether the model trained on 50,000 cards ultimately translates into real business value — it's too early to say. But if the question is "are these companies serious about AI," I think looking at what scenario data they hold and what business logic they've baked into their model designs gives you a pretty good answer.
A food delivery company training a trillion-parameter model isn't the strangest thing. The strangest thing is that you find it strange.