克雷西 明敏 发自 凹非寺 量子位 | 公众号 QbitAI 没想到啊，最新SOTA的开源大模型…… 来自一个送外卖 （ Waimai ） 的——有两个AI，确实不一样。 要知道，这可是一家"外卖公司"啊（手动狗头），做的模型都比Meta好了。 这个最新开源模型叫： Longcat-Flash-Chat ，美团第一个开源大模型，发布即开源，已经在海内外的技术圈子里火爆热议了。 一方面是因为成绩亮眼： 它在部分benchmark上，比如Agent工具调用、指令遵循的表现 超过DeepSeek-V3.1、Qwen3 MoE-2507 ，甚至比闭源的Claude4 Sonnet还要好。 编程能力也值得关注，在TerminalBench上， 和公认的"编程之王"Claude4 Sonnet不相上下 。 比如非常流行的小球氛围编程测试，LongCat编写的程序，运行起来效果是这样的： 另一方面是技术报告中透露出不少美团对于大模型的理解，包括DSMoE、MLA、动态计算、Infra等等。 我觉得这是中国大模型里最讲得详细的论文了，甚至超过Kimi、GLM，特别是在建模和infra方面。 而且不光是模型性能好，技术报告里还 ...

Core Viewpoint - The article discusses the launch of Meituan's open-source large model, Longcat-Flash-Chat, highlighting its impressive performance and technical innovations, which have sparked significant interest in the tech community both domestically and internationally [2][70]. Group 1: Model Performance - Longcat-Flash-Chat has outperformed several established models, including DeepSeek-V3.1 and Claude4 Sonnet, in various benchmarks, particularly in agent tool invocation and instruction adherence [3][18]. - The model's programming capabilities are noteworthy, showing comparable performance to Claude4 Sonnet in programming tasks [5]. - Longcat-Flash-Chat achieved a throughput improvement due to its unique architecture, which includes a "zero-computation expert" design, allowing it to dynamically activate parameters based on context [12][19]. Group 2: Technical Innovations - The model employs a dual design of "zero-computation experts" and Shortcut-connected MoE, which enhances training and inference throughput by allowing parallel execution of computations [12][16]. - Longcat-Flash-Chat has a total parameter count of 560 billion, which is lower than that of its competitors like DeepSeek-V3.1 and Kimi-K2, while still maintaining high performance [11][19]. - The model's training utilized over 20 trillion tokens in just 30 days, with a utilization rate of 98.48%, demonstrating its efficiency [19]. Group 3: Company Background and Strategy - Meituan's foray into large models is seen as a surprising development given its reputation as a food delivery company, but it has been building a foundation in AI through previous investments and projects [70][71]. - The establishment of the independent AI team GN06 and the launch of various AI applications indicate Meituan's commitment to integrating AI into its business model [73][74]. - Meituan's AI strategy focuses on practical applications, aiming to enhance employee efficiency and innovate existing products through AI technologies [87][85].

Core Viewpoint - AI programming tools are reshaping software development with a focus on "development democratization," evolving from simple code completion assistants to collaborative partners capable of understanding natural language requirements and generating runnable code frameworks [2] Group 1: Product Development and Features - Meituan launched its first AI Coding Agent product, NoCode, on June 10, 2023, aiming to establish its core competitiveness in the AI programming market [2] - The NoCode project started in October 2024 and was released in May 2023, with a focus on internal support and rapid product prototype delivery [3] - The AI Coding efficiency is complex to measure, with current observations focusing on AI-generated code's incremental proportion and adoption rate [2][3] Group 2: Model Optimization and Performance - The team optimized smaller models to balance performance and output quality, as larger models tend to have lower throughput speeds [4] - The self-generated code by NoCode indicates a low investment in development, with a small team achieving significant results [3][4] Group 3: User Experience and Target Audience - NoCode targets non-technical users, aiming to help them create functional products without extensive programming knowledge, while also being usable by technical users [6][7] - The product's design considers the needs of both novice users and experienced developers, focusing on creativity and continuous learning [7] Group 4: Future Directions and Challenges - The future of AI programming tools may shift from traditional IDE extensions to more autonomous agents capable of handling complex tasks [11] - The integration of various technologies and backend capabilities is essential for addressing complex product development challenges [10][12]

Core Insights - Meituan officially launched its first AI Coding Agent product, NoCode, on June 10, which is a conversational AI programming tool designed to automate coding tasks through natural language interactions [1][2] - NoCode integrates Meituan's self-developed LongCat model, which has a scale of hundreds of billions of parameters and performs at a leading level in code generation benchmarks [1] - The product has been widely applied in various development scenarios, showcasing its practicality and ability to simplify the development process for users without programming experience [2][3] Group 1 - NoCode allows users to create custom websites and software tools without any coding background, making it accessible for "0-basic" users [1] - The model is optimized for web front-end development, efficiently understanding and generating HTML, CSS, and JavaScript [1] - Meituan's R&D investment reached 5.8 billion yuan in Q1, a 15% year-on-year increase, indicating a strong commitment to AI development [1] Group 2 - The product emerged from an internal "hackathon" and was developed in just six months, highlighting the company's agile innovation process [2] - Meituan plans to launch additional AI application assistants within the year, further enhancing its AI capabilities [1][2] - The introduction of NoCode is expected to help small and medium-sized businesses reduce their reliance on SaaS services and outsourcing teams, strengthening the connection between the platform and merchants [2]

Core Insights - Meituan has officially launched its first AI Coding Agent product, NoCode, which allows users with no programming background to create websites and software tools through natural language interactions [1][2] - The NoCode tool integrates Meituan's self-developed LongCat model, which has a scale of hundreds of billions of parameters and performs comparably to leading models like GPT-4o and Claude-3.5-Sonnet in various benchmarks [1] - The product has been optimized for web front-end development, enabling efficient understanding and generation of HTML, CSS, and JavaScript code [1] Group 1 - NoCode was developed in just six months, originating from an internal "hackathon" event, and has already been widely used in various applications such as mini-games and ordering systems during its beta testing phase [2] - The official release of NoCode has further streamlined the development, testing, deployment, and bug-fixing processes, allowing users to complete complex tasks quickly without programming experience [2] - Approximately 50% of new code generated by Meituan employees is now produced by AI tools, with over 90% of engineering team members utilizing AI programming tools [2] Group 2 - The launch of NoCode aims to simplify programming tasks for small and medium-sized businesses, lowering the barriers to IT and digitalization [2] - Meituan plans to gradually open up more capabilities of its self-developed AI products in the future to promote technological inclusivity [2]