[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-muse-wonder-agentic-model":3,"topics-all":38,"news-related-9c8731f9-e170-4855-93a2-b1705abb61c3":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"9c8731f9-e170-4855-93a2-b1705abb61c3","Muse Wonder发布:Meta把宝押在个人智能体上","Meta 发布 Muse Wonder 与 Wonder Full,自称在个人及专业智能体基准上达到 SOTA。核心是 Active Memory 跨对话记住用户上下文,Active Focus 持续推进长任务。距推倒重建技术栈仅 14 个月,Full 版月费 50\u002F200 美元。","时隔十四个月推倒重建，Meta 的 AI 技术栈交出了第一份\"前沿级\"答卷：公司发布 Muse Wonder 与 Muse Wonder Full 两款模型，官方称其在个人与专业智能体任务的关键基准上达到 SOTA，并把它定位为通往\"个人超级智能\"的下一步（[Meta 官方博客](https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fmuse-wonder)）。\n\n## 这次真正的主角是\"记忆\"和\"专注\"\n\nMuse Wonder 的两个核心卖点都指向同一个方向——让智能体真正可用。**Active Memory** 指模型跨对话携带与用户相关的重要上下文，并在任务需要时主动检索细节；**Active Focus** 指模型面对长线目标时能持续工作：换路径、绕障碍、适应新信息，必要时主动向用户确认后再行动。配套评测也专门为此设计——用 FocusBench 量专注度，用 SWE-Marathon 测长时间软件项目的持续执行力，法律领域则引入 Harvey 的 Legal Agent Benchmark，要求模型基于原始文档产出法律工作成果。这些都不是传统刷分榜单，而是奔着\"数字员工能不能扛住真实工作\"去的。\n\n## 更强底座的隐性收益\n\n博客里最有信息量的是一段泛化分析：Meta 对比了 Muse Spark 1.3 与 Wonder Full 在\"训练目标内任务\"与\"陌生任务\"上的表现随推理算力扩展的曲线，发现 Wonder Full 在分布外任务上的表现超出旧趋势线的预测。官方将其归因于更强的底座模型——底座越强，后训练增益向陌生任务的迁移越充分。这其实是社区常说的\"大模型味\"（big model smell）的量化版本：榜单分数接近的两个模型，处理没见过的任务时差距可能很大，而个人智能体恰恰要面对无穷无尽的意外组合——目标、工具、约束的搭配没法在训练时全部预演。\n\n## 商业信号：前沿能力开始标价\n\nMuse Wonder 通过 Meta Model API 和 Muse Code 开放；更强的 Wonder Full 目前只面向 Muse Code 订阅者，据 CNBC 报道，订阅档位为每月 50 或 200 美元。第三方媒体 TechXplore 的表述是，Meta 声称新模型的表现与 OpenAI、Anthropic 的模型相当。训练侧的投入写在明面上：这批模型在 Meta 吉瓦级 Prometheus 数据中心完成训练。安全侧，官方称这是其\"对齐程度最高\"的模型，对 CBRN（化学、生物、放射、核）与网络滥用提供迄今最强的防护，同时按其 Superintelligence Scaling Framework 为防御用途保留前沿网络安全能力。\n\n## 所以呢\n\nMeta 把\"记住你、替你干活、持续干活\"做成了基准与产品双线发布，这比单纯的跑分更新更值得注意——智能体竞赛的评分维度正在从\"答得对\"转向\"扛得住\"。而对 Meta 自己，广告之外的新营收故事，现在有了月费 50 美元起步的注脚。下一次判断模型发布值不值得看，不妨先问一句：它的评测到底在测什么。\n","https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fmuse-wonder","245423b3-0e3c-47a8-9370-34e7a3b3988e",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d7e1c39d-19c8-4479-ae38-35227d024ec1","en","Meta Launches Muse Wonder for Personal AI Agents","Meta ships Muse Wonder and Wonder Full, claiming SOTA on agent benchmarks. Active Memory and Focus anchor it; Full tier costs $50-$200 monthly.","Fourteen months after a ground-up overhaul of its AI stack, Meta has delivered its first \"frontier-grade\" answer: Muse Wonder and Muse Wonder Full. The company claims state-of-the-art results on key benchmarks for personal and professional agents, positioning the series as its next step toward \"personal superintelligence\" ([Meta's official blog](https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fmuse-wonder)).\n\n## The real stars are memory and focus\n\nBoth headline capabilities of Muse Wonder point at the same goal — making agents genuinely usable. **Active Memory** means the model carries important context about you across conversations and actively retrieves the details relevant to the task at hand. **Active Focus** describes sustained effort toward a challenging goal: trying alternative approaches, working through blockers, adapting to new information, and asking for clarification or checking in before important actions. The evaluation suite was built for exactly this — FocusBench quantifies focus, SWE-Marathon tests sustained execution on long-running software projects, and the legal domain uses Harvey's Legal Agent Benchmark, which requires models to work from source documents and produce legal work products. These are not your traditional leaderboard runs; they target whether a \"digital worker\" can survive real work.\n\n## The hidden dividend of a stronger base\n\nThe most informative part of the blog is a generalization analysis. Meta compared Muse Spark 1.3 and Wonder Full on how performance scales with test-time compute, across tasks targeted during post-training (\"in-distribution\") versus tasks that were not (\"out-of-distribution\"). Wonder Full outperforms unfamiliar tasks beyond what the Muse Spark 1.3 trend would predict — an upward shift Meta attributes to larger, stronger base models: a stronger base lets post-training gains carry over more effectively to unfamiliar tasks. This is a quantified version of what the community calls \"big model smell\": two models with near-identical benchmark scores can diverge sharply on unseen tasks, and a personal agent faces endless unanticipated combinations of goals, tools, and constraints that no training run can fully rehearse.\n\n## The commercial signal: frontier capability gets a price tag\n\nMuse Wonder is available through the Meta Model API and Muse Code; the stronger Wonder Full is currently reserved for Muse Code subscribers — CNBC reports the plans run $50 or $200 per month. TechXplore's coverage notes Meta claims the new models perform at a level comparable to those of OpenAI and Anthropic. The training-side investment is stated plainly: the models were trained at Meta's gigawatt-scale Prometheus data center. On safety, Meta calls this its most aligned model family to date, with its strongest safeguards yet against CBRN (chemical, biological, radiological, nuclear) and cyber misuse, while preserving frontier cyber capabilities for defensive use under its Superintelligence Scaling Framework.\n\n## So what\n\nMeta has shipped \"remembers you, works for you, keeps working\" as both benchmark and product — a more meaningful shift than another leaderboard shuffle, because the agent race is being re-scored from \"answers correctly\" to \"holds up under load.\" For Meta itself, the story of revenue beyond advertising now has a footnote starting at $50 a month. The next time a model release crosses your feed, ask one question first: what is its evaluation actually measuring?\n","meta-muse-wonder-agentic-model","2026-10-12T23:10:00Z","2026-10-12T23:08:37.546823Z","2026-10-12T23:08:37.546837Z",true,"agent",52,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"27240196-3ac0-4af3-aec7-b0a4d3e7c990","智谱GLM-Legal法律Agent模型:Harvey榜前十最便宜","glm-legal-harvey-lab-cheapest","2026-10-13T15:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"5edbfca7-7c1a-4df3-8577-a0195408b827","RSI-Anything:世界模型陪跑,自我改进有了回滚键","rsi-anything-world-model-recursive-self-improvement","2026-10-12T21:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"8474959b-1569-45d6-afcc-34599f05b2fd","Memento 3:不动权重,LLM自我改进写进规则手册","memento-3-recursive-self-improvement-rulebook","2026-10-11T13:07:13+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"e40de2f9-9ee4-47f3-b1f1-7bf93a4870a9","一个动词翻转工具调用决策,LLM 内部向量现形","llm-tool-call-decision-vector","2026-10-09T13:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"c368ad9f-9308-4a0b-8f5c-3ae4601b48b9","D2K-Bench: 专家设计把 LLM 写 GPU 核提速 33.9%","d2k-bench-llm-gpu-kernel-design-guidance","2026-10-07T03:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"85f7f1c2-d896-436b-a915-37faed8776eb","你改主意了,模型没改:被拒需求也会带偏大模型","intent-eval-rejected-change-confusion","2026-10-06T17:15:00+00:00"]