[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fUdPPObfSLfZGs1F65BxouveB9LwhaTKq4yfI-rDBc8I":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1308,"美团发布并开源 LongCat-Flash-Chat，动态计算开启高效 AI 时代","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">9月1日，美团宣布 LongCat-Flash-Chat 正式发布，在Github、Hugging Face 平台开源，并同步上线官网。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Ffa29de51f5d04424b21ef51f4da3b272\u002FAA1LBK7S.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\" class=\"ql-align-center\">\u003Cp class=\"ql-align-center\">\u003Cspan class=\"ql-lineHeight-1-75\" style=\"color: rgb(187, 187, 187);\">▲美团发布并开源 LongCat-Flash-Chat（资料图）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">据悉，LongCat-Flash 采用创新性混合专家模型（Mixture-of-Experts， MoE）架构，总参数 560B，激活参数 18.6B-31.3B（平均 27B），实现了计算效率与性能的双重优化。根据多项基准测试综合评估，作为一款非思考型基础模型，LongCat-Flash-Chat 在仅激活少量参数的前提下，性能比肩当下领先的主流模型，尤其在智能体任务中具备突出优势。此外，因为面向推理效率的设计和创新，LongCat-Flash-Chat 具有明显更快的推理速度，更适合于耗时较长的复杂智能体应用。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002F76eed3d3a914438cb701f81431bc73dc\u002FAA1LBCV3.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-center\">\u003Cspan class=\"ql-lineHeight-1-75\" style=\"color: rgb(187, 187, 187);\">▲LongCat-Flash的基础测试性能（资料图）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">今年以来，美团 AI 进展频传，发布了 AI Coding Agent 工具 NoCode 、AI 经营决策助手袋鼠参谋、酒店经营的垂类 AI Agent 美团既白等多款 AI 应用。公司方面曾表示，其 AI 战略会建立在三个层面：AI at work、AI in products 以及 Building LLM，此次模型开源是其 Building LLM 进展的首度曝光。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">具体来看，LongCat-Flash 模型在架构层面引入“零计算专家（Zero-Computation Experts）”机制，总参数量 560B，每个token 依据上下文需求仅激活 18.6B-31.3B 参数，实现算力按需分配和高效利用。为控制总算力消耗，训练过程采用 PID 控制器实时微调专家偏置，将单 token 平均激活量稳定在约 27B。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">此外，LongCat-Flash 在层间铺设跨层通道，使MoE的通信和计算能很大程度上并行，极大提高了训练和推理效率。配合定制化的底层优化，LongCat-Flash 在 30 天内完成高效训练，并在 H800 上实现单用户 100+ tokens\u002Fs 的推理速度。LongCat-Flash 还对常用大模型组件和训练方式进行了改进，使用了超参迁移和模型层叠加的方式进行训练，并结合了多项策略保证训练稳定性，使得训练全程高效且顺利。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fb47e11fc423a40f29e8f4ff90028f336\u002FAA1LBAEQ.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-center\">\u003Cspan class=\"ql-lineHeight-1-75\" style=\"color: rgb(187, 187, 187);\">▲LongCat-Flash架构图（资料图）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">针对智能体（Agentic）能力，LongCat-Flash 自建了 Agentic 评测集指导数据策略，并在训练全流程进行了全面的优化，包括使用多智能体方法生成多样化高质量的轨迹数据等，实现了优异的智能体能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">通过算法和工程层面的联合设计，LongCat-Flash在理论上的成本和速度都大幅领先行业同等规模、甚至规模更小的模型；通过系统优化，LongCat-Flash 在 H800 上达成了 100 token\u002Fs的生成速度，在保持极致生成速度的同时，输出成本低至5元\u002F百万 token。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】金融界财经 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002F%E6%8A%80%E6%9C%AF\u002F%E6%8A%80%E6%9C%AF%E5%85%AC%E5%8F%B8\u002F%E7%BE%8E%E5%9B%A2%E5%8F%91%E5%B8%83%E5%B9%B6%E5%BC%80%E6%BA%90-longcat-flash-chat-%E5%8A%A8%E6%80%81%E8%AE%A1%E7%AE%97%E5%BC%80%E5%90%AF%E9%AB%98%E6%95%88-ai-%E6%97%B6%E4%BB%A3\u002Far-AA1LBOrB?ocid\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FWlbBK\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002F8b72dfe058f84ff3b72e7b22bf5adda8\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fthumbs\u002F8b72dfe058f84ff3b72e7b22bf5adda8\u002FAI领域.jpg",0,1,44,"2025-09-01 16:51",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A7bd9f172-33b3-434b-a0b1-13d049ca688d%3A0.wav?Expires=1756721188&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=wq7XTriC5Z6WbC6BJq90f2UrxpA%3D",5621830,"7bd9f172-33b3-434b-a0b1-13d049ca688d","2025-09-01 16:47","Meituan releases and opens sources LongCat-Flash-Chat, dynamic computing starts the efficient AI era","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">On September 1st, Meituan announced that LongCat-Flash-Chat was officially released, open-sourced on Github and Hugging Face platforms, and also launched on its official website.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Ffa29de51f5d04424b21ef51f4da3b272\u002FAA1LBK7S.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\" class=\"ql-align-center\">\u003Cp class=\"ql-align-center\">\u003Cspan class=\"ql-lineHeight-1-75\" style=\"color: rgb(187, 187, 187);\">▲Meituan releases and open sources LongCat-Flash-Chat (reference picture)\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">According to the information, LongCat-Flash adopts an innovative mixture-of-experts (MoE) architecture, with a total of 560B parameters, activating 18.6B-31.3B parameters (average 27B), achieving dual optimization of computing efficiency and performance. According to comprehensive evaluation of multiple benchmark tests, as a non-thinking base model, LongCat-Flash-Chat performs comparable to current leading mainstream models with only a small number of activated parameters, especially showing outstanding advantages in agent tasks. In addition, due to its design for inference efficiency and innovation, LongCat-Flash-Chat has significantly faster inference speed, making it more suitable for complex agent applications that take longer time.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002F76eed3d3a914438cb701f81431bc73dc\u002FAA1LBCV3.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-center\">\u003Cspan class=\"ql-lineHeight-1-75\" style=\"color: rgb(187, 187, 187);\">▲Basic test performance of LongCat-Flash (reference picture)\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This year, Meituan's AI progress has been frequent, releasing multiple AI applications such as the AI Coding Agent tool NoCode, the AI business decision assistant Kangaroo Consultant, and the vertical AI Agent Meituan Jibai for hotel operations. The company stated that its AI strategy will be based on three levels: AI at work, AI in products, and Building LLM. This model open-source is the first exposure of its progress in Building LLM.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Specifically, the LongCat-Flash model introduces a \"Zero-Computation Experts\" mechanism at the architecture level, with a total parameter count of 560B, and each token activates 18.6B-31.3B parameters according to context needs, achieving on-demand allocation and efficient use of computing power. To control total computing power consumption, the training process uses a PID controller to dynamically adjust expert bias in real-time, stabilizing the average activation per token at around 27B.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In addition, LongCat-Flash lays cross-layer channels between layers, allowing MoE communication and computation to be largely parallel, greatly improving training and inference efficiency. Combined with customized low-level optimization, LongCat-Flash completes efficient training within 30 days and achieves a reasoning speed of 100+ tokens\u002Fs per user on H800. LongCat-Flash also improves common large model components and training methods, using superparameter transfer and model layer stacking for training, and combining multiple strategies to ensure training stability, making the entire training process efficient and smooth.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F09\u002Fb47e11fc423a40f29e8f4ff90028f336\u002FAA1LBAEQ.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-center\">\u003Cspan class=\"ql-lineHeight-1-75\" style=\"color: rgb(187, 187, 187);\">▲Architecture diagram of LongCat-Flash (reference picture)\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Regarding agentic capabilities, LongCat-Flash has built its own Agentic evaluation set to guide data strategy, and has comprehensively optimized throughout the training process, including using multi-agent methods to generate diverse high-quality trajectory data, achieving excellent agentic capabilities.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Through joint design at the algorithm and engineering levels, LongCat-Flash has significantly led the industry in terms of cost and speed for models of the same scale, and even smaller scales; through system optimization, LongCat-Flash achieves a generation speed of 100 tokens\u002Fs on H800, and maintains ultra-fast generation speed while keeping output costs as low as 5 yuan per million tokens.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">[News source] FinanceNet \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002F%E6%8A%80%E6%9C%AF\u002F%E6%8A%80%E6%9C%AF%E5%85%AC%E5%8F%B8\u002F%E7%BE%8E%E5%9B%A2%E5%8F%91%E5%B8%83%E5%B9%B6%E5%BC%80%E6%BA%90-longcat-flash-chat-%E5%8A%A8%E6%80%81%E8%AE%A1%E7%AE%97%E5%BC%80%E5%90%AF%E9%AB%98%E6%95%88-ai-%E6%97%B6%E4%BB%A3\u002Far-AA1LBOrB?ocid\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FWlbBK\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is reprinted by this site to provide readers with more information and news. The content involved does not constitute investment or consumption advice. If there are any questions about the facts of the article, please verify with the relevant parties. The views expressed in the article are not the views of this site, and are for reference only.)\u003C\u002Fspan>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A1c442b40-501d-409b-9248-3b696f458d4e%3A0.wav?Expires=1774838480&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=pzHSkXDTO67HCdnO4wcTcQGbgCU%3D","1c442b40-501d-409b-9248-3b696f458d4e",6915284]