Sakana Fugu 是否只是个 LLM Router

{"type":"doc","content":[{"type":"image","attrs":{"id":"aafe9fd6-1592-45e5-a731-f01e8d283b64","src":"https://developer.qcloudimg.com/http-save/audit-1972990/f60890800cf9c0870acb6e4663a62f50.png","extension":"png","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":990,"aspectRatio":"2.216495","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"a567e09b-443c-49d5-aac0-ba680e6759a2","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"背景"}]},{"type":"paragraph","attrs":{"id":"6c879583-dda2-4466-9086-4e232c86e862","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Sakana Fugu 来自日本 AI 公司 Sakana AI。Sakana AI 总部位于东京，由 David Ha、Ren Ito 和 Llion Jones 创立，其中 Llion Jones 是 Transformer 论文《Attention Is All You Need》的共同作者之一。"}]},{"type":"paragraph","attrs":{"id":"143ffe96-9bf7-43a6-8d6a-9526327ce86f","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Sakana AI 一直强调「自然启发」「进化算法」「集体智能」。Fugu 正是这条路线的产品化结果：它不追求只训练一个更大的单体模型，而是让多个强模型像团队一样协作。"}]},{"type":"heading","attrs":{"id":"9bfaed3f-a41c-442a-9618-ceb412cb4bbd","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Fugu是什么"}]},{"type":"blockquote","attrs":{"id":"1f5a492d-90b5-4192-b0c2-25b9876a9d5b","textAlign":"inherit","isHoverDragHandle":false},"content":[{"type":"paragraph","attrs":{"id":"61e0c719-ccfc-4b94-a656-e82e6c2633a6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"A Multi-Agent System, Delivered as One Model"}]}]},{"type":"paragraph","attrs":{"id":"51652a3d-7b95-40bc-8d95-e027f6ee5939","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Fugu 是一个「以单一模型 API 形式交付的多智能体系统」。"}]},{"type":"paragraph","attrs":{"id":"009756a4-8ad1-46ca-aae5-721097d84957","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"用户调用它时，看起来只是调用一个普通模型，例如 "},{"type":"text","marks":[{"type":"code"}],"text":"fugu"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":" 或 "},{"type":"text","marks":[{"type":"code"}],"text":"fugu-ultra"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"。但在背后，Fugu 会动态选择多个模型，让它们分别承担推理、执行、验证、总结等角色，最后再合成一个结果。"}]},{"type":"paragraph","attrs":{"id":"82e47995-2ca0-42bf-86e1-1f63365f7426","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"它的关键不是「又一个大模型」，而是「一个会组织模型协作的模型」。"}]},{"type":"image","attrs":{"id":"406cd302-4d40-42fe-9b90-9e0867108a7a","src":"https://developer.qcloudimg.com/http-save/audit-1972990/e71e67158aaf1db0a1e3b1814a7ef77a.png","extension":"png","align":"center","alt":"","showAlt":false,"href":"","boxShadow":"","width":990,"aspectRatio":"1.828331","status":"success","showText":true,"isPercentage":false,"percentage":0,"isHoverDragHandle":false}},{"type":"paragraph","attrs":{"id":"578308a3-198b-4082-9fd6-cc844377dea6","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false}},{"type":"heading","attrs":{"id":"f7a5e0f1-3fee-41d7-8284-0024c801a562","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"时间节点"}]},{"type":"paragraph","attrs":{"id":"1f336605-d01e-4fd9-bc30-4dd78e41915b","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Sakana AI 在 2026 年 6 月正式发布 Fugu。相关技术基础来自 ICLR 2026 论文，包括 TRINITY 和 Conductor。"}]},{"type":"paragraph","attrs":{"id":"968b1116-a93a-4c6f-af14-9ed055624171","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"TRINITY 研究如何让多个 LLM 以 Thinker、Worker、Verifier 等角色协作。Conductor 则进一步训练一个协调模型，让它学会为不同任务动态设计协作结构。"}]},{"type":"heading","attrs":{"id":"433e5c62-361a-4082-a80e-38581cc49efb","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"适用场景"}]},{"type":"paragraph","attrs":{"id":"94124c6a-ace9-431e-91f4-ad3c5473a12c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Fugu 最适合复杂、多步骤、需要验证的任务，例如："}]},{"type":"bulletList","attrs":{"id":"c75a34e1-3bd5-4d53-b72c-5e211dab7056","isHoverDragHandle":false},"content":[{"type":"listItem","attrs":{"id":"2e07bf13-66d3-4dba-858c-f622c1c1c8ea"},"content":[{"type":"paragraph","attrs":{"id":"6165be78-160e-4a09-8c17-bcf291abab21","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"代码生成与代码审查"}]}]},{"type":"listItem","attrs":{"id":"6ccf8f2b-fe7c-47fd-ade7-8f3071145194"},"content":[{"type":"paragraph","attrs":{"id":"159fe2a2-022e-42eb-94ad-4227c66cc871","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"软件工程 Agent"}]}]},{"type":"listItem","attrs":{"id":"985bba38-c5b9-4520-8f55-c4444b31a127"},"content":[{"type":"paragraph","attrs":{"id":"e4a69b08-9025-4e09-a476-ea1d618aec06","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"数学和逻辑推理"}]}]},{"type":"listItem","attrs":{"id":"16ef16c7-893a-4a2e-9966-678c5e6eae72"},"content":[{"type":"paragraph","attrs":{"id":"67cb61f6-b9bc-4de5-a489-bd6c6328276c","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"科研论文分析"}]}]},{"type":"listItem","attrs":{"id":"6591c3ed-342f-44ee-a194-3e77bcf0ad1f"},"content":[{"type":"paragraph","attrs":{"id":"325ab4c6-eb64-46cf-968c-21211bed860a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"长文档研究"}]}]},{"type":"listItem","attrs":{"id":"4e496d7e-9830-4d85-a0d2-1c1c2ce29edc"},"content":[{"type":"paragraph","attrs":{"id":"1acf7655-dc3d-4789-9587-03e125792abe","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"专利、法律、金融等复杂材料分析"}]}]},{"type":"listItem","attrs":{"id":"9ac785c8-2c0b-4b8a-b117-cc0db700ba62"},"content":[{"type":"paragraph","attrs":{"id":"1f495ea7-eaa7-4eb8-8e51-8f829814105d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"需要多模型交叉验证的高风险任务"}]}]}]},{"type":"paragraph","attrs":{"id":"170c0b0c-1d58-46a7-b4e7-7efccafdfe7d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"如果只是普通聊天或简单问答，Fugu 的优势未必明显。但当任务需要拆解、执行、检查、修正时，多智能体协作就可能比单模型更稳定。"}]},{"type":"heading","attrs":{"id":"89d2f81b-048b-46eb-8e99-53498eb87f81","textAlign":"inherit","indent":0,"level":3,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"价格"}]},{"type":"paragraph","attrs":{"id":"b74b10a2-c96a-4181-8404-c053d713367a","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Fugu Ultra 的固定计费为"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}},{"type":"bold"}],"text":"每100 万 token 输入 5 美元、输出 30 美元、缓存输入 0.50 美元；当上下文超过 272K 时，输入 10 美元、输出 45 美元、缓存输入 1 美元"},{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"。普通 Fugu 的计费更像「动态路由账单」：如果只激活 1 个 Agent，按底层模型标准价格计费；如果多个 Agent 参与，官方称不会叠加多个模型费用，而是按参与模型中的最高层级收取单一费率。"}]},{"type":"heading","attrs":{"id":"7ec5c7aa-11fe-4db4-91d1-abb1bfd5f5e9","textAlign":"inherit","indent":0,"level":2,"isHoverDragHandle":false},"content":[{"type":"text","text":"思考"}]},{"type":"paragraph","attrs":{"id":"ef67e3d5-9423-453a-b117-5b9ddba52956","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"Fugu 代表了 AI 模型竞争的新方向。"}]},{"type":"paragraph","attrs":{"id":"90d219ec-afa8-4e8c-bc27-a11e2fa98357","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"过去的竞争重点是「谁的单模型更强」。Fugu 提出的方向是：未来的 AI 能力不一定来自一个最大模型，而可能来自一个更会调度模型的系统。"}]},{"type":"paragraph","attrs":{"id":"3fc98f11-bf1a-4ec1-8438-695ee4c83e49","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","marks":[{"type":"textStyle","attrs":{"color":"","background":""}}],"text":"这对企业尤其重要。真实业务往往不是一次问答，而是连续流程：理解需求、查资料、写代码、跑测试、发现错误、修正方案、生成报告。单模型容易在长流程里漂移，而多智能体系统可以通过分工和互相验证提升稳定性。"}]},{"type":"paragraph","attrs":{"id":"d542f675-c1ee-43ce-823e-b229450f3d8d","textAlign":"inherit","indent":0,"color":null,"background":null,"isHoverDragHandle":false},"content":[{"type":"text","text":"参考来源："},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://sakana.ai/fugu/","target":"_blank","rel":"noopener noreferrer nofollow","class":null}}],"text":"Sakana Fugu"},{"type":"text","text":"、"},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://arxiv.org/abs/2512.04695","target":"_blank","rel":"noopener noreferrer nofollow","class":null}}],"text":"TRINITY 论文"},{"type":"text","text":"、"},{"type":"text","marks":[{"type":"link","attrs":{"href":"https://arxiv.org/abs/2512.04388","target":"_blank","rel":"noopener noreferrer nofollow","class":null}}],"text":"Conductor 论文"}]}]}","createTime":1782112859,"ext":{"closeTextLink":0,"comment_ban":0,"description":"","focusRead":0},"favNum":0,"html":"","isOriginal":0,"likeNum":0,