The video explains how Meituan, a Chinese food delivery giant, has quickly become a major force in AI research through its LongCat lab, releasing advanced open-source language and multimodal models that rival or surpass those from tech giants like Meta. Meituan’s innovations in efficiency, transparency, and technical depth—especially in language and video generation—make it a standout player in the global AI landscape.
The video discusses how Meituan, a major Chinese food delivery company often compared to DoorDash, has rapidly become a significant player in the AI space, outperforming even tech giants like Meta in some areas. Despite its reputation as a food delivery service, Meituan has a long history of technical innovation and research, publishing numerous papers and technical blogs since 2013. In September 2025, Meituan launched its AI research lab called LongCat, which has since released a series of impressive open-source AI models and benchmarks in just a few months, including language models, audio and video encoders, and multimodal systems.
LongCat’s debut language model, FlashChat, stands out for its technical depth and innovation. The accompanying 36-page technical report details everything from pre-training to distributed training strategies, achieving a remarkably low cost of $0.50 per million tokens. The model introduces a novel context-aware dynamic computation mechanism, allowing the system to allocate more computational resources to harder-to-predict tokens by leveraging a mixture-of-experts (MoE) architecture. This approach not only improves efficiency but also maintains high performance, making FlashChat highly competitive with other state-of-the-art models.
Beyond language models, LongCat has also made significant strides in generative video modeling. Their video generation model introduces a 3D block sparse attention mechanism, which drastically reduces computational requirements while maintaining near-lossless quality. The lab has also unified text-to-video, image-to-video, and video continuation tasks into a single input framework, a notable advancement over current industry standards that often rely on less efficient cross-attention mechanisms. These innovations are particularly valuable given the typically closed-source nature of top-tier video generation models.
LongCat’s research extends into multimodal and omnimodal models as well. Their omnimodal model, built on top of FlashChat, integrates visual and audio inputs and outputs, achieving performance on par with leading private models like Qwen 3 Omni, GT40, and Gemini 2.5. Importantly, LongCat is unique in openly sharing detailed infrastructure optimizations for these complex models, including techniques like modality-decoupled parallelism and chunk-based modality bridges. This transparency and willingness to share knowledge reflect Meituan’s longstanding culture of open research and technical blogging.
Overall, the video highlights Meituan’s rapid and impressive progress in AI research, emphasizing their focus on optimization and efficiency. The company’s research outputs are described as a gold mine for anyone interested in advanced AI infrastructure, making Meituan a key player to watch in 2026. The presenter also promotes their own educational website for learning about language models, aiming to make these complex topics accessible to a broader audience.