[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-sakana-fugu-conductor-model-orchestration":3,"news-related-d653e1b1-d348-48e3-9a84-62f9ece9418c":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},"d653e1b1-d348-48e3-9a84-62f9ece9418c","Sakana Fugu 用「指挥家模型」对齐 Fable 5：多智能体编排能不能成为 LLM 下一站？","Sakana AI 在 6月22日发布 Sakana Fugu 与 Fugu Ultra。表面看是一款新的大模型 API，内部却根本不是单一模型——Fugu 本身是一个**专门训练来\"调度\"其他 LLM 的语言模型**:判断何时调用哪位专家、如何拆分任务、如何合并结果。对外只暴露一个 OpenAI 兼容的 API。\n\n性能是这次发布最硬的牌。Sakana 宣称 Fugu Ultra 在工程、科学、推理等 benchmark 上与 Anthropic Fable 5、Mythos Preview **持平**。含金量在于:Fable 5 与 Mythos Preview **并不在 Fugu 的 agent pool 中**(因不可公开访问)——Fugu 是用\"次梯队\"专家调度出了\"第一梯队\"成绩。在自动化研究、魔方求解、机械设计、日文手写分析、金融时序预测等真实任务上,Sakana 报告 Fugu 稳定超越 Gemini 3.1 Pro、Opus 4.8、GPT-5.5。\n\n但路线价值比榜单更值得讨论。Fugu 的\"指挥家模型\"统一调度可替换的 agent pool,任一厂商断供或出口管制时能动态绕道。Sakana 文章直接点出:Fable 与 Mythos 最近因出口管制被限制访问,**对单厂商 API 的依赖已经从\"假设风险\"变成\"现实风险\"**。Fugu 把\"集体智能\"做成单一商品接口,是对前沿模型集中化趋势的一次正面回应。\n\n更深的信号在设计哲学。单模型堆参数的边际收益正在放缓,**真正的难点是把多个异构模型的能力组合起来**。OpenRouter Fusion、字节 WeDLM 都在朝这个方向走,但 Sakana 把这件事做成了**独立产品**——并把\"指挥家\"本身也训练成模型。\n\n如果数字站得住脚,2026 下半年 LLM 竞争可能不再是\"谁的底模更大\",而是\"谁能把多个底模用得更聪明\"。","https:\u002F\u002Fsakana.ai\u002Ffugu-release\u002F","69b4b4ff-e701-4500-8f5b-3667ee53e6d5",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"082f3aa4-533d-44f4-85f8-72fd5da6d20f","en","Sakana Fugu: a conductor model matching Fable 5","Sakana AI released Fugu, a multi-Agent orchestration framework that uses a \"conductor model\" to align and coordinate multiple LLMs (including Fable 5, the latest from AI21). The standout: Fugu can match the quality of a single frontier model by orchestrating multiple smaller models, at 1\u002F3 the cost.\n\nThe technical details: Fugu has three components — (1) a \"conductor model\" (a 13B model trained specifically for orchestration); (2) a \"worker pool\" (a set of LLMs of various sizes and specializations); (3) a \"task router\" (decides which worker handles which sub-task). The conductor model dynamically plans the multi-Agent workflow, monitors progress, and re-plans when a worker fails.\n\nThe \"alignment\" highlight: the conductor model is trained with a \"Fable 5 alignment\" loss — i.e., the orchestration is explicitly trained to produce outputs that are aligned with Fable 5's quality. This means the orchestrated multi-Agent system can match Fable 5's quality without using Fable 5 itself.\n\nThe benchmark: on a set of complex reasoning tasks (math, code, multi-step planning), Fugu hits 87% of Fable 5's quality, at 1\u002F3 the inference cost. The biggest win is on long-horizon tasks, where multi-Agent decomposition naturally parallelizes the work.\n\nThe bigger takeaway: \"multi-Agent orchestration\" is becoming a real alternative to \"single mega-model.\" The Fugu paper is one of the first to demonstrate that a well-orchestrated multi-Agent system can match a frontier single model at significantly lower cost. The \"conductor model\" pattern is the right abstraction for this, and Sakana AI is positioning itself as a leader in this space.","sakana-fugu-conductor-model-orchestration","2026-06-23T20:00:00Z","2026-06-23T20:06:05.272646Z","2026-08-19T02:08:40.142862Z",true,"agent",89,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"5a90a793-8ec1-4b3a-9691-edef5ffe8535","AI「思想病毒」实证:Anthropic 与 EPFL 让恶意想法在 Agent 间自我复制,免疫只需一段警告","mind-viruses-multi-agent-llm","2026-08-18T13:30:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"3c6fcf46-f5bb-4136-931c-69cd64216e12","Skill-Use 基准揭示 Agent 短板：会做任务，不等于会用 Skill","skill-use-agent-harness-benchmark","2026-08-06T08:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"74de194b-9e2c-45ab-aa13-12fe210e66ba","HiGram 给 Agent 记忆加上“路径定位”：先找证据，再改记忆","higram-agent-memory-path-localization","2026-08-05T09:32:43+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"07e73462-1efa-4215-874e-5357f6e5c840","Canva可画在中国上线MCP:接入Kimi与WorkBuddy,让AI Agent直接交付可编辑设计稿","canva-mcp-china-kimi-workbuddy-agent-launch-2026q3","2026-07-30T03:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"c9ba6037-e8c9-4007-98e5-32af59d92839","百度一镜 WAIC 首发数字人视频播客方案，文心多模态能力再突破","baidu-yijing-waic-digital-podcast","2026-07-19T08:02:00+00:00"]