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].

Group 1 - Huawei has launched a new product, referred to as the "computer version of Moutai," with a starting price of 23,999 yuan, sparking discussions about its high pricing and potential risks associated with its large foldable screen [1] - The National Cybersecurity and Information Security Information Reporting Center has identified 35 mobile applications, including several popular AI apps, for illegally collecting and using personal information [5] - The shopping mall "Pang Dou Lai" has changed its name to "Ying Dou Lai" after facing legal pressure from the well-known retail company "Pang Dong Lai" due to the similarity in names [7] Group 2 - Zhong Shanshan, at the Nongfu Spring shareholders' meeting, stated that while he does not oppose OEM (Original Equipment Manufacturer) practices, all of Nongfu Spring's products are currently not suitable for outsourcing due to their high dependency on water sources and complex production systems [10] - Meng Yutong has hinted at a potential live-streaming collaboration with her former boss, Dong Mingzhu, after a two-year hiatus, with both parties expressing a willingness to reconnect [13] - Vogue's parent company Condé Nast has appointed Sherry Lang, former head of Tmall Luxury, as the new General Manager for Vogue China, marking a shift towards leaders with diverse backgrounds in luxury fashion, e-commerce, and digital technology [15]