Meituan Trained a 1.6-Trillion-Parameter Model on Domestic Silicon. The Clever Part Is What It Built It For.

On November 2, Meituan released LongCat-2.5, a natively multimodal model with 1.6 trillion total parameters and roughly 48 billion activated per token.
Two details in the release are being under-reported outside China. The first is that training ran entirely on domestic Chinese compute, and Meituan calls it the largest training task completed on that hardware. The second is that Meituan's chief executive has publicly said the company will not become a token factory.
Put those together and LongCat-2.5 reads less like a bid for the frontier and more like a very large industrial tool built for one company's business.
The architecture, briefly
The visual side is the interesting engineering. LongCat-2.5 represents visual signals as multi-level discrete tokens rather than passing continuous embeddings to the language backbone. Following the method used in LongCat-Next, native-resolution features from a semantic-and-aligned encoder go through residual vector quantisation, which turns an image into something the transformer can process as tokens, in the same way it processes words.
The result is a single token space rather than a bolt-on vision encoder, which is why Meituan can claim native multimodality and point at tasks like interactive web pages, 3D display pages and creative design as things the model handles rather than things it was adapted to handle.
The training scale is documented too. Mid-training used sequences up to 1 million tokens, with trillions of tokens of cumulative training. Multi-card long-context work is where memory capacity and inter-card communication become the bottleneck, so Meituan optimised task partitioning around supernodes, keeping the bulk of parameter and data traffic on the fast interconnect inside a node rather than crossing between them.
The efficiency numbers are concrete. On long-context supervised fine-tuning, single-card training throughput improved 86% over a distributed data parallel baseline, which Meituan frames as one card doing the work of about 1.86. For inference, the model uses W4A4 quantisation, pushing both expert weights and activations to four bits, tuned for the characteristics of domestic accelerators. Against a W8A8 INT8 baseline, throughput on input processing and per-token generation rose about 40%, first-token latency fell roughly 16%, and per-token generation time fell about 20%, with downstream task performance held roughly level.
That last set of numbers is the actual achievement. Getting a trillion-parameter model to train and serve on hardware that is not the market leader is an engineering problem about memory bandwidth and communication, and Meituan's June 2026 release of LongCat-2.0, the first trillion-parameter model fully trained and served on domestic compute, was the warm-up.
Benchmarks, with the usual caveat
Meituan reports that LongCat-2.5 lands between Claude Opus 4.8 and Claude Opus 5 across a range of mainstream benchmarks, and exceeds Opus 5 on some.
These are vendor-reported comparisons against a competitor's models, on a benchmark selection the vendor made. Treat them as evidence that the model is competitive rather than as a ranking. The more useful signal is that a Chinese local-services company is now shipping a model in the same conversation as Anthropic's frontier line, and publishing the training-efficiency work alongside it.
What it was built for
The stated design goal is long-horizon agent tasks. Meituan's description covers adding long-horizon trajectory data to long-sequence training, running large-scale long-horizon reinforcement learning, and converting real agent failure trajectories into reproducible, verifiable training tasks. That last practice is the one more labs should copy: turning your own agent's failures into a dataset is cheaper than scraping more of the internet.
Then the commercial logic. Meituan runs food delivery, in-store services and travel across thousands of Chinese cities. Its merchants are small operators with thin margins. From that position, a better chatbot ranks well below software that reads a merchant's channel data, tells them where they are losing money, and then does something about it.
The example Meituan published is worth quoting in detail. In a multi-channel brand operating analysis, the model built the channel profit and loss statement itself, then identified that paid GMV had grown 23.2% while channel profit had fallen 1.7%. It connected that to clearance economics, repeat purchase behaviour and inventory risk, and proposed differentiated resource allocation.
That is a real analytical chain. What stands out is that the model chose to compute the margin ratio alongside the growth number, which is the thing an experienced operator does and a dashboard does not.
Meituan has wrapped this into a product called Miaoshou, an AI operating workbench for merchants. It understands industry metrics and brand rules, and within merchant authorisation it calls tools and systems to complete tasks. Meituan says it is already serving chains including Guming, Xushengji and Yihetang, across brand analysis, store management and customer acquisition.
The strategic argument
Wang Xing's line from the August earnings call is the clearest statement any Chinese platform chief executive has made about model strategy. AI's long-term value, he said, is whether it genuinely enters real workflows, solves real problems and becomes stable infrastructure. Meituan will not be a token factory. Its models and products exist to support the main business.
Read that against the past year of frontier lab behaviour and the difference is stark. OpenAI, Anthropic and Google are building models to sell tokens and subscriptions. Meituan is building a model to improve the margin of a milk tea chain.
Both are legitimate. They produce very different organisations. A token factory optimises for benchmark leadership and API price. A captive model optimises for the specific reasoning an operator needs and for the cost of running it at the scale of your own transaction volume, which is why the W4A4 work matters more here than an arena score would.
Serving cost is the quiet headline
The four-bit quantisation deserves a separate note because it addresses the constraint that actually limits deployment outside a lab.

A 1.6-trillion-parameter model with 48 billion parameters active per token does not need the whole model in fast memory for any single token, but it does need the active experts to be reachable quickly, and at four bits the weights occupy a fraction of the space they would at sixteen. Meituan reports roughly 40% higher throughput on both input processing and per-token generation against an INT8 baseline, with first-token latency down about 16% and per-token time down about 20%, and downstream quality holding.
For a company serving its own transaction volume, that is the difference between an AI feature that costs more than the margin it protects and one that does not. It is also the kind of work that does not produce a leaderboard entry, which is why it rarely gets reported.
The open questions
Meituan did not publish weights for LongCat-2.5, so the domestic-compute claim cannot be independently examined, and the model is available through longcat.ai and an API rather than as an open release. The efficiency gains are measured against baselines Meituan defined.
There is also the usual tension in captive models: the better it gets at Meituan's business, the less useful it may be outside it. A model tuned on local-services profit and loss may be excellent at restaurant unit economics and unremarkable at legal drafting.
For everyone outside China, the takeaway is the shift in the domestic-silicon constraint from a compliance problem to an engineering programme with published throughput numbers. Once a company can train 1.6 trillion parameters and serve them at four-bit precision on hardware it can actually buy, the practical gap between frontier and follower narrows to the things money cannot fix: data, taste and distribution.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.