[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-kimi-k2-5-agent-swarm-100-subagents":3,"news-related-a9e4bd12-171c-47d3-8ecd-2c532aac9daf":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"a9e4bd12-171c-47d3-8ecd-2c532aac9daf","Kimi K2.5解锁Agent Swarm：百个AI子代理并行协作重塑大规模任务效率","月之暗面（Moonshot AI）旗下Kimi K2.5近日在多模态模型基础上解锁了一项关键新能力——Agent Swarm。这一多代理协作框架允许模型同时调度最多100个AI子代理，每个子代理独立执行搜索、生成、分析、信息整理等任务，以并行方式完成复杂的长周期工作流。\n\n从单代理到多代理，代表着架构范式的跨越。传统AI模型以单代理执行为主，面对大规模任务时往往面临效率瓶颈——即使模型推理能力再强，单线执行也有天然的速度上限。Agent Swarm的核心创新在于将任务进行智能切分，由主代理（Kimi K2.5）统一协调，根据任务性质动态分配给子代理池，实现真正的并行处理。根据官方披露，Agent Swarm在大型研究、长篇内容创作、批量下载等场景中，可将任务执行时间缩短至传统单代理模式的1\u002F4.5，效率提升显著。\n\nKimi K2.5本身是原生多模态模型，支持文本、图像、视频的输入理解，以及前端代码的高保真生成。Agent Swarm的加入使得模型从“能做什么”扩展到“能协同完成什么”——这在工程层面代表了一种从模型能力到系统能力的跃升。更值得关注的是，这种多代理编排并非依赖外部调度框架，而是模型本身通过指令理解自主完成子代理的生成、分配与结果整合。这意味着模型需要具备强大的任务分解能力与上下文管理能力——而这恰恰是长上下文窗口（Kimi K2.5支持262K tokens）的核心价值所在。\n\n从GPT-5.5的多代理工具调用，到DeepSeek V4的长上下文优化，再到Kimi K2.5的Agent Swarm，多代理协作正在成为头部模型厂商竞争的新焦点。相比单纯追求Benchmark分数，多代理系统更贴近真实工作场景的需求，也更容易转化为生产力工具。对于开发者而言，Kimi K2.5的Agent Swarm意味着可以基于单一API构建复杂AI工作流，无需额外集成第三方代理框架。而对行业来说，这代表着开源模型正在快速补齐与闭源模型在Agent能力上的差距，竞争将进一步加剧。\n\nKimi K2.5 Agent Swarm目前已在Kimi平台开放，支持Web、App、API及Kimi Code多端访问。","https:\u002F\u002Fwww.kimi.com\u002Fai-models\u002Fkimi-k2-5","0ec8f614-42c7-4256-8591-209e1e39eb6b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"472553c3-7c26-4701-a7c3-f8aab382760b","en","Kimi K2.5 unlocks Agent Swarm: 100 sub-agents in parallel","Moonshot AI's Kimi K2.5 has recently unlocked a key new capability on top of its multimodal model — Agent Swarm. This multi-agent collaboration framework allows the model to dispatch up to 100 AI sub-agents simultaneously, each independently executing tasks like search, generation, analysis, and information organization, completing complex long-horizon workflows in parallel.\n\n**From single-agent to multi-agent: an architectural paradigm leap**\n\nTraditional AI models rely on single-agent execution, which faces efficiency bottlenecks when handling large-scale tasks — even with the strongest model reasoning capability, single-thread execution has a natural speed ceiling. Agent Swarm's core innovation is the intelligent task decomposition: the main agent (Kimi K2.5) coordinates uniformly, dynamically dispatching tasks to the sub-agent pool based on task nature, achieving true parallel processing. According to official disclosure, in scenarios like large research, long-form content creation, and batch downloads, Agent Swarm can shorten task execution time to 1\u002F4.5 of traditional single-agent mode, a significant efficiency boost.\n\nKimi K2.5 itself is a native multimodal model, supporting text, image, and video input understanding, as well as high-fidelity front-end code generation. Agent Swarm's addition extends the model from \"what it can do\" to \"what it can collaboratively accomplish\" — at the engineering level, this represents a leap from model capability to system capability. More notably, this multi-agent orchestration doesn't depend on external scheduling frameworks, but the model itself autonomously handles sub-agent generation, dispatch, and result integration via instruction understanding. This means the model needs strong task-decomposition ability and context-management ability — exactly the core value of long-context windows (Kimi K2.5 supports 262K tokens).\n\nFrom GPT-5.5's multi-agent tool calling, to DeepSeek V4's long-context optimization, to Kimi K2.5's Agent Swarm, multi-agent collaboration is becoming a new focus of competition among top model vendors. Compared to simply pursuing benchmark scores, multi-agent systems are closer to real work-scenario needs and easier to convert into productivity tools. For developers, Kimi K2.5's Agent Swarm means complex AI workflows can be built on a single API without integrating third-party agent frameworks. For the industry, this means open-source models are rapidly closing the gap with closed-source models in agent capability, and competition will further intensify.\n\nKimi K2.5 Agent Swarm is now open on the Kimi platform, supporting Web, App, API, and Kimi Code multi-end access.","kimi-k2-5-agent-swarm-100-subagents","2026-05-14T13:00:00Z","2026-05-14T13:08:03.016655Z","2026-08-19T02:08:40.142862Z",true,"agent",305,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"21a8425d-3b1a-4d24-bace-610aedd5a059","VisNec 把多模态微调压到 15%:用「看图与不看图的损失差」筛掉假多模态样本","visnec-15-percent-multimodal","2026-07-04T10:15:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"58d2e247-e1d7-4325-8e90-602480fae550","微信AI团队ICASSP 2026获奖：从视觉冗余切入，让VLM在边缘设备真正跑起来","wechat-icassp-2026-vlm-edge-best-paper","2026-05-19T02:30:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"ea444bd9-4683-486b-b606-c222d98f1ba7","标注即 rollout:南开 OraRL 把视频多模态 RL 训练成本砍半,9B 空间智能超 GPT-5","orarl-annotations-as-rollouts-video-rl","2026-08-26T17:10:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"34edaffc-6b5c-4df1-9e2f-d864cada6063","Gemini 走进 K-12 课堂：Google 把「上下文」塞进每个作业","gemini-classroom-k12-contextualized-prompts","2026-08-07T02:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"d7b6d14d-7257-4794-b92f-31956bbc7eae","原生多模态 vs 后训练加压:国产头部基模两条路线的工程账","native-multimodal-vs-posttraining-2026","2026-08-05T00:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"2d29aa3d-317c-4126-a6f7-2c9c2c3b6f93","Kimi K3与DeepSeek V4之间,隔着原生多模态的时间差","kimi-k3-deepseek-v4-native-multimodal-divergence","2026-08-04T08:02:10+00:00"]