# Chinese DoorDash Is Making Better LLMs Than Meta Source: https://www.youtube.com/watch?v=9GWOksNjFpY Recap page: https://rapidrecap.app/video/9GWOksNjFpY Generated: 2026-01-28T22:04:49.409+00:00 --- ## Quick Overview Meituan researchers are significantly advancing AI by releasing numerous open-source models and technical papers, demonstrating strong performance across various benchmarks, particularly in agentic capabilities, which surpasses the output of closed models like those from Meta. **Key Points:** - Meituan researchers released LongCat-Flash, a 560-billion parameter Mixture-of-Experts (MoE) model, trained on over 20 trillion tokens in just 30 days for $0.70 per million tokens. - The LongCat team has published over 700 technical blogs and papers, including LongCat-Flash, LongCat-Audio-Codec, LongCat-Video, and LongCat-Flash-Omni, demonstrating consistent research output since 2019. - LongCat-Flash achieves high throughput and low inference latency, exceeding 100 tokens per second (TPS) on H800 hardware during inference. - The LongCat-Flash-Omni model is state-of-the-art for open-source multi-modal tasks, outperforming models like Qwen3-Omni, GPT-4o, and Gemini 2.5 Flash on several OmniBench, Audio, Video, and Image benchmarks. - Meituan's Meituan (DoorDash equivalent in China) is projected to have $55.3 billion in estimated 2025 revenue, nearly four times the revenue of DoorDash. - The LongCat models leverage techniques like Context-Aware Dynamical Computation Mechanism and MoE++ architecture to maintain high performance while reducing compute cost. - The video highlights the comprehensive nature of Meituan's AI research, covering everything from fundamental LLM concepts to advanced multi-modal architectures. ![Screenshot at 0:05: The video begins by displaying the logos of major Chinese tech companies like Xiaomi \(electronics\), Xiaohongshu \(social media\), and Meituan \(food delivery\), setting the stage for discussing AI advancements from these firms.](https://ss.rapidrecap.app/screens/9GWOksNjFpY/00-00-05.jpg) **Context:** This video examines the significant and rapidly growing contributions of Meituan's AI research team, known as LongCat, to the field of large language models (LLMs) and multi-modal AI. The video contrasts Meituan's open-source output with closed models from major tech companies like Meta, suggesting that Meituan's research is setting new standards in efficiency and capability, especially in agentic tasks, challenging the dominance of US-based tech giants. ## Detailed Analysis The video argues that Chinese DoorDash competitor, Meituan, through its research group LongCat, is producing superior LLMs compared to Meta's offerings. Meituan's research output is prolific, with over 700 technical blogs and papers published since 2019, covering topics from LLM fundamentals to advanced multi-modal architectures. The LongCat team released LongCat-Flash, a 560B parameter MoE model trained on over 20 trillion tokens in only 30 days at a low cost of $0.70 per million tokens, achieving over 100 tokens per second (TPS) inference speed on H800 GPUs. The architecture employs novel techniques like the Context-Aware Dynamical Computation Mechanism and MoE++ routing to dynamically activate only necessary experts, leading to efficient compute and cheap inference. Furthermore, the LongCat-Flash-Omni model demonstrates state-of-the-art performance against competitors like DeepSeek-V3.1, Qwen3-MoE-2507, and Gemini 2.5 Flash across numerous general, reasoning, and agentic benchmarks, often leading in performance while utilizing fewer activated parameters (27B vs. 37B for DeepSeek-V3.1). The video also highlights Meituan's strong business standing, projecting $55.3 billion in 2025 revenue, second only to Uber, far surpassing DoorDash. The commitment of Meituan's researchers to open-sourcing their advancements—including models like LongCat-Flash and LongCat-Flash-Omni—is presented as a key differentiator from closed-source competitors, providing valuable resources to the wider AI community. ### Chinese Tech Giants in AI - Xiaomi produces phones and is making LLMs - Xiaohongshu is a Chinese Pinterest - Meituan is a Chinese DoorDash company ### LongCat Research Output - Released LongCat-Flash (560B MoE) in 30 days, training on >20T tokens - Published LongCat-Flash-Omni (multi-modal) - Published LongCat-Video, Audio-Codec, and Flash-Thinking models ### LongCat-Flash Technical Achievements - Achieved 100+ TPS inference on H800 at $0.7/million tokens - Used expert-specific bias term for dynamic computation allocation - Reused the router mechanism from DeepSeek for efficiency ### Performance Benchmarks - LongCat-Flash beats DeepSeek-V3.1 and others in agentic tool use and instruction following - LongCat-Flash-Omni outperforms Qwen3-Omni, GPT-4o, and Gemini 2.5 Flash in Omni, Audio, Video, and Image benchmarks ### Meituan Business Context - Estimated 2025 revenue of ~$55.3 Billion (Rank 2 globally for food delivery, behind Uber) - Founder Wang Xing started the company in May 2010, pivoting through group-buying, movie ticketing, and finally food delivery in 2013. ### Key Innovations - Context-Aware Dynamical Computation Mechanism for expert activation - Zero-computation experts (Z) to keep parameter count stable - Modality-decoupled parallelism (MDP) and chunk-based ModalityBridge for infrastructure efficiency ### Learning Resources - IntuitiveAI.academy offers a structured path from LLM fundamentals to cutting-edge topics like LoRA and MoE, with a New Year discount code NYNM for 50% off yearly plans. ![Screenshot at 0:01: An animated scene showing shapes on a conveyor belt being sorted, representing the concept of efficient computation or routing.](https://ss.rapidrecap.app/screens/9GWOksNjFpY/00-00-01.jpg) ![Screenshot at 0:05: Logos of major Chinese tech companies \(Xiaomi, Xiaohongshu, Meituan\) are displayed, setting the context for the video's focus on Chinese AI advancements.](https://ss.rapidrecap.app/screens/9GWOksNjFpY/00-00-05.jpg) ![Screenshot at 0:06: Title slide for the LongCat MiMo paper, highlighting the focus on unlocking reasoning potential from pretraining to posttraining.](https://ss.rapidrecap.app/screens/9GWOksNjFpY/00-00-06.jpg) ![Screenshot at 0:24: An animation illustrating Meituan's \(yellow square logo\) model \(LongCat, green M logo\) replacing a Meituan block on a production line, symbolizing the development of LLMs by a non-traditional tech company.](https://ss.rapidrecap.app/screens/9GWOksNjFpY/00-00-24.jpg) ![Screenshot at 0:35: A slide showing a detailed list of statistics about the 44 people in Meta's Superintelligence team, used as a comparative baseline for talent distribution.](https://ss.rapidrecap.app/screens/9GWOksNjFpY/00-00-35.jpg)