[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-muse-spark-1-3-meta-agent-release":3,"topics-all":38,"news-related-d400c0db-49df-4cc6-a87e-87b709f59fea":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},"d400c0db-49df-4cc6-a87e-87b709f59fea","Muse Spark 1.3 发布:卡住会向用户求助的 Agent,工具调用少 20%、token 省 25%","Meta 发布 Muse Spark 1.3:提示含糊会追问、卡住会向用户求助、高风险操作前先确认;编码场景工具调用约少 20%、token 约省 25%,并增强抗提示注入能力。已上线 Muse Code 与 Meta Model API,官方预告将放出开放权重版本。","9 月 2 日,Meta Superintelligence Labs 在官方博客发布了 Muse Spark 系列的新版本 Muse Spark 1.3,主打 agentic 与编码两类任务的实用性升级,并以 max reasoning 档位上线 Muse Code 与 Meta Model API,入口在 dev.meta.ai([官方公告](https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-spark-1-3))。\n\n## 从「会干活」到「会配合」\n\n这版最大的变化不是跑分,而是 Agent 行为模式。官方描述的场景是:面对一个开放式目标,Muse Spark 1.3 会自己用工具在杂乱、相互矛盾的来源里生成上下文,主动补上计划里的缺口,并记住已经学到的东西,最后产出一份可交付的成果——全程在一条长线程里同时推进多个工作流。\n\n更值得注意的是协作姿态。提示含糊时,它会**主动追问**而不是硬猜;卡住时会**向用户求助**;要执行有后果的操作前会**先确认**。长任务里它还能适应你的偏好——要么频繁汇报进度,要么安静地在后台干活。这些听起来像产品细节,但恰恰是 Agent 从 demo 走向生产环境最缺的那一环:知道自己不知道什么。官方也明确训练了模型对自身能力边界的感知,遇到障碍时不幻觉编造结果。\n\n## 编码侧:少说话,多办事\n\n相对 1.2,Muse Spark 1.3 在训练中加入了更多长程编码任务。Meta 工程师的内部对比显示,它明显更快、更省:工具调用约**少 20%**,token 消耗约**少 25%**,不需要的轮次更少、输出更不啰嗦,整体代码风格也更干净。\n\n对一个每天要跑成千上万次 Agent 会话的团队来说,这两组数字比任何榜单都实在——token 就是账单,工具调用就是延迟。\n\n## 安全与生态位\n\n安全侧官方强调两点:更强的对抗鲁棒性(抗对抗输入与提示注入),以及在复杂 agentic 任务上对「不可逆操作」的校准更准——什么该做、什么该先问,分寸感更好。\n\n生态位上,Muse Spark 1.2 与 Muse Code 在 8 月初发布后经过了一个多月的大规模采用,1.3 是在这批真实反馈上迭代出来的。第三方工具链也在跟进:开源 Agent 框架 OpenClaw 的 2026.9.2 版本已宣布支持 GPT-6 Astra 与 Muse Spark 1.3,并把子 Agent 编队(swarm)设为默认开启。\n\n官方在结尾预告了路线图:**更大的模型、Muse Spark 开放权重版本**等。Meta 用「personal superintelligence」描述这版的目标——口号先不论,开放权重这条对开发者社区是真金白银的承诺,值得盯着后续落地。\n\n对普通开发者的启示很简单:Agent 模型的竞争焦点正在从「谁更聪明」转向「谁更省、谁更知道分寸」。当能力差距缩小,效率和边界感就成了新的分水岭。","https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-spark-1-3","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},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"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},"dd6af4b6-a23a-4783-b849-a5a4deab40f5","en","Muse Spark 1.3 arrives: the agent that asks for help","Meta ships Muse Spark 1.3: agent asks for help when stuck, ~20% fewer tool calls, ~25% fewer tokens, resists prompt injection. Open weights teased.","Meta released Muse Spark 1.3 on September 2, focusing on agentic and coding usability rather than raw benchmark scores. Available with max reasoning on Muse Code and the Meta Model API at dev.meta.ai ([official announcement](https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-spark-1-3)).\n\n## From capable to collaborative\n\nThe biggest change is not scores but agent behavior. Given an open-ended objective, Muse Spark 1.3 uses tools to generate its own context across messy, conflicting sources, proactively corrects gaps in its plan, and tracks what it has learned to produce a final deliverable — while juggling multiple workflows in a single long thread.\n\nThe collaborative posture is more notable. When prompts are ambiguous, it asks clarifying questions instead of guessing; it invokes help from the user when stuck; it confirms before consequential actions. On long tasks it adapts to user preferences, either providing frequent updates or working silently in the background. Meta also explicitly trained the model's awareness of its own capabilities and limitations, so it avoids hallucinating outcomes when it hits hurdles.\n\n## Coding: fewer calls, fewer tokens\n\nRelative to 1.2, Muse Spark 1.3 was trained on more long-horizon coding tasks. In comparisons by Meta engineers, it proved significantly faster and more efficient: roughly 20% fewer tool calls and 25% fewer tokens, with fewer unnecessary turns, less verbose output, and a cleaner coding style.\n\nFor teams running thousands of agent sessions a day, these two numbers matter more than any leaderboard — tokens are the bill, and tool calls are the latency.\n\n## Safety and ecosystem\n\nOn safety, Meta highlights stronger adversarial robustness (better resistance to adversarial inputs and prompt injections) and better calibration on what counts as irreversible actions in complex agentic tasks.\n\nEcosystem-wise, 1.3 builds on weeks of broad adoption of Muse Code and the Meta Model API since the 1.2 release in early August. Third-party tooling is following: OpenClaw 2026.9.2 has added support for both GPT-6 Astra and Muse Spark 1.3, with sub-agent swarms on by default.\n\nThe official roadmap teases bigger models and a Muse Spark open-weights release. Meta frames this release as progress toward \"personal superintelligence\" — whatever the slogan means, the open-weights promise is real value for the developer community.\n\nThe takeaway is simple: agent-model competition is shifting from \"who is smarter\" to \"who is cheaper to run and knows its limits.\" As capability gaps narrow, efficiency and judgment become the new dividing line.","muse-spark-1-3-meta-agent-release","2026-09-06T15:12:00Z","2026-09-06T15:10:02.657432Z","2026-09-06T15:10:02.657442Z",true,"agent",222,[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},"8ebbcd9c-31ee-4baa-b395-b104bd87c8e1","Kimi K2.8 Preview 把 K3 的百万上下文下放给免费档：月之暗面的「过日子」模型登场","kimi-k2-8-preview-coding","2026-09-17T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"01a6593b-449e-493e-ac45-33c23c9211ba","SWE-2 距 Fable 5.1 一分:2.8T 开源底座后训练,成本砍 64%","cognition-swe-2-kimi-k3-pareto","2026-09-11T21:08:02+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"63c30bcd-3ffc-47c5-bd74-c2a9ed8f7c94","DeepSeek Harness 预览版开源:Agent 被拆成可插拔的插件栈,模型只负责想、Harness 负责做事","deepseek-harness-plugin-stack","2026-09-05T06:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"c77381d9-29ba-45ff-89df-855d11f90de2","Terminal-Universe:Qwen 把旧轨迹反向重建为 3.73 万个环境,27B 微调双基准 +11.9\u002F+13.8 分","terminal-universe-trajectory-environments","2026-09-04T17:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"14a7f5ab-e270-461c-b862-4bde139e463f","HarnessDev 基准:让 LLM 自建 Agent Harness,代码领域仍输人类工程师","harnessdev-llm-selfbuilt-agent-harness","2026-09-03T19:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"71dde565-b87a-4470-83a5-0ad7a4ee1787","IBM Granite 4.2 开源:原生思维链做成开关,30B 拿下 SWE Bench Pro","ibm-granite-4-2-native-reasoning-agents","2026-08-28T14:00:00+00:00"]