Vibeleaderboard
Meituan

ai lab

Meituan

Meituan matters because it connects foundation-model research to one of the world's largest physical-service networks. LongCat also provides unusually concrete evidence that a major sparse model can be trained and served on a large non-Nvidia cluster, while its open weights let developers test whether the reported coding and agentic results survive outside Meituan's evaluation stack.2,9,7,11

Official site ↗Models in the Atlas →

Profile

Overview

A local-services platform becomes an AI operator

Meituan is a Beijing technology company founded in 2010 by Wang Xing. It began with a group-buying website, merged with Dianping in 2015, and developed into a large marketplace for food delivery, travel, retail, and local services. That operating business matters to its AI work because the company controls recommendation, search, dispatch, merchant, mapping, and fulfillment systems that run at national scale.1,2

Machine learning begins inside the marketplace

Machine learning was part of Meituan before it became a public model publisher. The company applied prediction and optimization to delivery dispatch, estimated arrival times, demand, recommendations, and autonomous delivery. Its 2025 annual report says LongCat became the main foundation for internal AI applications, including consumer assistants that can recommend restaurants and hotels and carry out transactions inside Meituan's services.2,9

LongCat grows into a model family

LongCat became a visible open model program in 2025. LongCat-Flash used zero-computation experts and shortcut-connected mixture-of-experts layers to vary active computation and overlap communication with computation. Later releases extended the family into reasoning, video avatars, and a multimodal system that represents text, images, and audio as discrete tokens handled by one autoregressive backbone.3,4,5,6

LongCat 2.0 tests domestic training at scale

LongCat 2.0 moved the program into a different systems class in mid-2026. Meituan announced the 1.6-trillion-parameter sparse model on June 30 and published weights and inference materials under the MIT license in July. The model activates about 48 billion parameters per token and supports a one-million-token context window. Reuters and the South China Morning Post reported Meituan's claim that training and inference ran across more than 50,000 Chinese-made accelerators, but neither publication independently audited the hardware run, and broad third-party capability testing remains thinner than the company's benchmark table.7,8,9,10,11

Notable contributions

  1. 01Adaptive sparse computation in LongCat-FlashLongCat-Flash documented zero-computation experts and shortcut-connected mixture-of-experts layers as a practical recipe for varying token-level compute while improving communication overlap. The contribution is the published LongCat design, not mixture-of-experts modeling itself.3
  2. 02One token-prediction system across modalitiesLongCat-Next converts language, image, and audio inputs into discrete tokens and trains one autoregressive backbone to model them together. This gives Meituan a technical line beyond text-only commerce assistants.5
  3. 03Trillion-parameter training on domestic acceleratorsLongCat 2.0 paired public weights with a systems account of pretraining and inference on a cluster of more than 50,000 Chinese-made accelerators. Independent reporting confirms that Meituan made the claim and released the model, while the exact hardware composition and systems results remain company-reported.9,10,7