[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-meta-muse-spark-1-1":3,"news-related-73545307-3c18-48ab-aac9-5a4c7fc96b53":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},"73545307-3c18-48ab-aac9-5a4c7fc96b53","Meta Muse Spark 1.1：1M 上下文，Computer Use 自主决策","7月9日,Meta Superintelligence Labs(MSL)放出 Muse Spark 1.1,这是 4 月 Muse Spark 的首个主版本升级,定位从\"多模态推理模型\"正式切换到\"agentic foundation\":同一份权重里塞进 Main + Subagent 编排、零样本 MCP、自适应 Computer Use、视觉编程,以及 OpenAI 兼容的 Meta Model API 公测。\n\n最值得关注的不是任何单点指标,而是三件事被拧到一起:\n\n第一,1M token 上下文 + active compaction。1.1 能主动检索早期的操作轨迹、按重要性压缩噪声,只保留后续步骤真正需要的中间状态。这直接命中长程 Agent 最痛的地方——会话中段开始\"失忆\"导致 refactor 翻车。Meta 把上下文管理从工程技巧拉到模型内生能力。\n\n第二,真正的多 Agent 编排。Main agent 收集上下文→做计划→分发 parallel subagents;subagent 知道工具边界,卡住时 escalate 回主线程。这个\"角色分配 + 上下汇报\"的协议,过去是各家 harness 自己写,现在 Meta 直接写进了模型策略里。\n\n第三,Computer Use 从\"GUI 自动化\"扩到\"脚本 vs 点击\"二选一。模型会在每一步判断:写脚本更快,就直接 exec;点 UI 更简单,就点;每步还会 batch 出一组动作。这把 Codex Computer Use、Claude Computer Use 拉到的同一类工作流,从\"点点点\"推进到\"模型自己挑工具\"。\n\n落地层面,Meta 同步开了 Meta Model API(OpenAI 兼容公测),首批伙伴 Replit、Cline、Box 都已经接入做长上下文 + 工具调用的端到端跑通。配合 7 月 7 日的 Muse Image,Spark 1.1 实际上把\"感知(看图\u002F看视频)→ 推理(多 Agent)→ 执行(Computer Use + Coding)\"拼成了一条完整闭环。\n\n行业层面,这是 7 月 9 日 Frontier 栈的第四块拼图(同天 Grok 4.5、GPT-5.6、Ollama \\$88M):大厂都在把\"harness-ready 的模型 + 兼容 API\"做成新一代分发单元,纯聊天模型开始让位于\"agentic runtime\"。\n\n一个保留意见:Meta 自报的 Meta Internal Coding Bench 显著优于原版且\"competitive with leading alternatives\",但工具使用型 Agent 的第三方独立评测,往往是 real world 和 sandbox 之间最大的鸿沟——值得在自己 repo 上跑一遍再做生产决策。","https:\u002F\u002Fai.meta.com\u002Fblog\u002Fintroducing-muse-spark-meta-model-api\u002F","a1f0bda7-5035-4317-b63b-72693539d2e3",[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},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"c47d29b7-5d21-4214-8c8e-73783eea7b97","en","Meta Muse Spark 1.1: 1M context, autonomous Computer Use","On July 9, Meta Superintelligence Labs (MSL) released Muse Spark 1.1, the first major version upgrade of Muse Spark since April, with positioning formally shifting from \"multimodal reasoning model\" to \"agentic foundation\": the same weights now pack Main + Subagent orchestration, zero-shot MCP, adaptive Computer Use, visual programming, and the OpenAI-compatible Meta Model API public beta. The most noteworthy things aren't any single metric, but the fact that three things are now wound together: First, 1M token context + active compaction. 1.1 can actively retrieve earlier operation traces, compress noise by importance, and only keep the intermediate state that's truly needed for subsequent steps. This directly hits the long-horizon Agent's most painful place — \"amnesia\" in the middle of a conversation causes refactor to fail. Meta is pulling context management from engineering tricks to a model-internal capability. Second, true multi-Agent orchestration. The main agent collects context → makes a plan → dispatches parallel subagents; subagents know the tool boundary and escalate back to the main thread when stuck. This \"role assignment + upward reporting\" protocol was previously written by each vendor's own harness; now Meta has written it directly into the model policy. Third, Computer Use extends from \"GUI automation\" to \"script vs click\" either\u002For. The model decides at each step: if writing a script is faster, it just execs; if clicking the UI is simpler, it clicks; each step also batches out a set of actions. This pushes the same kind of workflow that Codex Computer Use and Claude Computer Use landed from \"pointing and clicking\" to \"model picks its own tool\". On the landing side, Meta simultaneously opened the Meta Model API (OpenAI-compatible public beta), with first partners Replit, Cline, and Box already integrated for end-to-end runs of long-context + tool-calling. Combined with the July 7 Muse Image, Spark 1.1 has actually stitched \"perception (see images \u002F see video) → reasoning (multi-Agent) → execution (Computer Use + Coding)\" into a complete closed loop. On the industry side, this is the fourth piece of the July 9 frontier-stack puzzle (same day as Grok 4.5, GPT-5.6, and Ollama's $88M): the big labs are turning \"harness-ready model + compatible API\" into the new distribution unit, and pure chat models are starting to give way to \"agentic runtime\". One reservation: Meta's self-reported Meta Internal Coding Bench significantly outperforms the original and is \"competitive with leading alternatives\", but third-party independent evaluations of tool-use Agents are often the biggest gap between real-world and sandbox — worth running on your own repo before making production decisions.","meta-muse-spark-1-1","2026-07-09T16:00:00Z","2026-07-09T16:09:56.865981Z","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},"c329665c-c1fd-4af4-ad28-a1fbf9a8ede7","Qwen3.7-Flash 上线：多模态 Agent 从演示走向 CI 与搜索一线","qwen-3-7-flash","2026-07-28T02:02:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"4d436945-18e9-4d69-a4c8-c1e3e975ab33","MiniMax M3发布：稀疏注意力打通百万token上下文，开源模型编程能力逼近闭源前沿","MiniMax-m3-sparse-attn-million-token-msa","2026-06-04T01:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"4c7f5330-3aff-458a-9ef5-f04cc5585703","微信视觉团队开源 WeMM 嵌入模型:2B 反超 8B 前基线,9B 达 MMEB-v2 80.6","wemm-embedding-wechat-multimodal","2026-08-26T21:07:30+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"e75069c6-f15c-4ff9-8b11-404d705442e8","Upstage Solar Pro 4:把「agent 跑得稳」做成新一代闭源模型卖点","upstage-solar-pro-4-agent-reliability-closed-llm","2026-08-25T03:00:00+00:00"]