[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-79c1684f-f61d-4799-b3d0-6450c4ad10e8":3},{"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},"79c1684f-f61d-4799-b3d0-6450c4ad10e8","Muse Glimmer:Meta 把 30B 「常驻本地的智能体」开源,把 Agent 拉到笔记本里 7×24 跑","Meta Superintelligence Labs 8 月 10 日把 30B 参数的 Muse Glimmer 以 Apache 2.0 协议开源,主打「always-on 本地 agent」场景,4-bit 量化 \u003C20GB 配 DFlash 投机解码,RTX 5090 上解码速度提至 3.1×,Ollama\u002Fllama.cpp\u002FExecuTorch 当日集成。","## 一、为什么 Meta 这次要单点 30B?\n\n很多人看到「30B」第一反应是「为什么不是 70B、不是 405B?」答案藏在 Muse Glimmer 的训练配方里:\n\n- **预训练(Pre-Training)**:用 Muse Spark 的输出做 logit 蒸馏,数据配比与教师模型接近。换句话说,Muse Glimmer 不是「从零开始学语言」,而是「从大老师那儿学推理与工具调用」。\n- **中训练(Mid-Training)**:换成更长上下文、更高 agent 密度、带丰富推理轨迹的数据,搭配 organic 数据。\n- **后训练(Post-Training)**:SFT + on-policy 蒸馏 + RL,覆盖通用、推理、代码、agentic 四个域。\n\n这条「蒸馏教师 → 中训练加 agent 味 → 后训练精调」的链路,意味着 30B 这个体量不是 Meta 妥协的结果,而是**为了单卡本地推理所精挑细选的甜点**。再大一点就塞不进 24\u002F32 GB 显存,再小一点就撑不住长程工具调用。\n\n## 二、量化 + DFlash 投机解码:把「慢」变成「实时」\n\n跑过 30B 模型的人都知道,光「能跑」和「能日常用」之间还差一道坎——首 token 延迟与解码速度。Meta 在这一步做了两件具体的事:\n\n1. **K-Quant 量化到 ~4-bit,语言模型压到 20 GB 以内**。剩下的显存空间留给 KV cache、perception encoder(图像理解)、以及 DFlash 草稿模型。整张模型、视觉模块、加速器同时跑在一张 24 GB 或 32 GB 的消费卡上,这是 Meta 给出的「能塞得下」的本地 agent 方案。\n2. **DFlash 投机解码**,Meta 把一个小型 drafter 网络挂在主模型旁边,一次提出一整块 token,主模型并行验证。官方给出的加速比:\n   - **RTX 5090:3.1×**\n   - **M5 Max:1.8×**\n   - **M4 Max:1.5×**\n\n3.1× 这个数字,本质上把 30B 模型的「对话感」拉到了和 7B 差不多的水准,这是「本地 agent 真能日常用」的临界值。\n\n## 三、Agent 能力到底在哪儿\n\nMeta 给的 benchmark 列表值得仔细看:DeepSearch QA、MCP-Atlas、τ-Bench、SWE-Bench。**这四个全部是 full-task agent 基准,不是单纯的知识问答**。具体能做什么:\n\n- **长程执行**:跨多轮、多步完成任务\n- **可靠 tool use**:复杂 schema 的函数调用,中途不崩\n- **多步推理**:长链规划,不掉线\n- **失败恢复**:工具调用失败时诊断 + 重试,而不是直接 halt\n- **多模态输入**:通过 dedicated perception encoder 吃 interleaved 文本 + 图像,可以解读截图、图表、文档\n- **可控 effort**:支持不同推理强度,让用户在质量与速度之间挑\n\n特别值得一提的是 **Scaffold 兼容**:Meta 明确写了 Muse Glimmer 适配 OpenClaw 等 agent 编排模式——这意味着如果你的工程栈已经在用 agent 编排框架(比如 OpenClaw 这类),Muse Glimmer 是 plug-in,不用改 scaffold。\n\n## 四、这步棋背后是 Meta「开源 + 闭源」的双轨\n\n把 Muse Glimmer(开源 30B 本地 agent)和几天前刚发的 Muse Spark 1.2(闭源)放一起看,Meta 的策略就清楚了:\n\n- **Muse Spark \u002F Spark 1.2**(闭源):打前沿能力、企业 API、agent 平台\n- **Muse Glimmer**(开源 30B):打本地、edge、personal agent、研究社区\n\n这是 Meta 一直强调的开源 AI 战略叙事的具体落地:**让开发者、研究者、甚至发烧友都能在自家电脑上跑出一个 24\u002F7 不掉线的 AI 助手**。\n\n合作伙伴名单也透出 Meta 的野心:Ollama、LM Studio、Unsloth、llama.cpp、ExecuTorch、MLX、vLLM、SGLang、Together AI、Fireworks AI、OpenRouter——**从本地到云,部署路径全覆盖**。硬件侧则拉上了 AMD、Arm、Dell、Intel、NVIDIA。\n\n## 五、所以呢?给读者的「所以」\n\nMuse Glimmer 不是一个「更大的 Llama」,它是 Meta 在「**AI 真正跑进个人设备**」这条路上的一次集中押注。30B 配 4-bit 量化 + 投机解码这个组合,如果后续开源社区把 llama.cpp \u002F MLX \u002F ExecuTorch 的集成做好,**日历管理、文件整理、代码助手、个人 RAG 检索**这些场景,可能会第一次出现「不联网也能用、跑得动、跑得起」的消费级本地 agent。\n\n对国内大模型生态的启示也很直接:**「够用 + 跑得动」比单纯「更大」更有商业价值**。Meta 这条「30B 量化 + 投机解码 + 蒸馏链路」的组合,给出了一个清晰的天花板——**想覆盖 24\u002F7 本地 agent 体验,30B 可能是当前这个工程节点上的甜点参数**。下一个值得盯的变量,是 DFlash 之后的下一代投机解码方案能不能把这个甜点继续往下推。","https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-glimmer-open-agentic-model","a1f0bda7-5035-4317-b63b-72693539d2e3",[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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"e061fe17-5e11-4b1f-96d7-ea56fe5db7f9","en","Muse Glimmer: Meta Open-Sources a 30B 'Always-On' Agent That Lives on Your Laptop","On August 10 2026, Meta Superintelligence Labs released the 30B-parameter Muse Glimmer under Apache 2.0, targeting always-on local agent workflows. 4-bit quantization brings it under 20 GB, and DFlash speculative decoding hits 3.1x speedup on RTX 5090, with day-one integrations across Ollama, llama.cpp, ExecuTorch, MLX, and the usual cloud partners.","# Muse Glimmer: Meta Open-Sources a 30B \"Always-On\" Agent That Lives on Your Laptop\n\n**On August 10 2026, Meta Superintelligence Labs released the weights of Muse Glimmer under Apache 2.0 on Hugging Face, going straight after the \"AI agent that runs 7×24 on a personal computer\" scenario nobody has actually shipped yet.** Thirty billion parameters, under 20 GB after quantization, and a target hardware envelope of a single consumer GPU on a laptop or desktop — this is the first time \"local agent\" has crossed from demo to a real engineering milestone.\n\n## 1. Why does Meta single out 30B?\n\nMany people will ask: why 30B, not 70B or 405B? The answer is in the training recipe.\n\n- **Pre-Training:** Muse Glimmer is trained on the outputs of Muse Spark using logit distillation, with a data mix close to the teacher. In other words, the model does not \"learn language from scratch\" — it learns reasoning and tool use from a much larger teacher.\n- **Mid-Training:** the data mix shifts to longer context, higher agent density, and richer reasoning traces, alongside organic data.\n- **Post-Training:** SFT combined with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains.\n\nThis \"distill-from-teacher → mid-train on agent-heavy data → fine-tune\" pipeline is not Meta settling for 30B. It is **a deliberately chosen sweet spot for single-card local inference**. Anything bigger will not fit in 24\u002F32 GB of VRAM; anything smaller will not sustain long-horizon tool use.\n\n## 2. Quantization + DFlash speculative decoding: turning \"slow\" into \"real-time\"\n\nAnyone who has actually run a 30B model knows the gap between \"it can run\" and \"I can use it daily\" comes down to time-to-first-token and decode speed. Meta does two concrete things here.\n\n1. **K-Quant to roughly 4-bit, compressing the language model under 20 GB.** The remaining VRAM headroom goes to KV cache, the perception encoder for image understanding, and the DFlash drafter model. The main model, vision module, and accelerator all run together on a single 24 GB or 32 GB consumer card. Meta's proposal is the \"fits\" local-agent setup.\n2. **DFlash speculative decoding.** Meta attaches a small drafter network to the main model. The drafter proposes a full block of tokens at once, and the main model verifies them in parallel. The official speedups:\n   - **RTX 5090: 3.1×**\n   - **M5 Max: 1.8×**\n   - **M4 Max: 1.5×**\n\nThe 3.1× figure essentially pulls a 30B model's conversational feel into the same range as a 7B. That is the threshold for \"a local agent I can actually use every day.\"\n\n## 3. Where the agent capability actually lives\n\nThe benchmark list from Meta is worth reading carefully: DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench. **All four are full-task agent benchmarks, not plain knowledge Q&A.** The capabilities:\n\n- **Long-horizon execution:** completing tasks across many turns and steps.\n- **Reliable tool use:** complex-schema function calls that survive long workflows.\n- **Multi-step reasoning:** coherent long-chain planning that does not fall apart.\n- **Failure recovery:** when a tool call fails or returns an unexpected result, the model diagnoses and retries rather than halting.\n- **Multimodal input:** a dedicated perception encoder accepts interleaved text and images, so agents can interpret screenshots, charts, and documents alongside conversation.\n- **Controllable effort:** the model supports different reasoning strengths, letting users trade quality for speed.\n\nWorth calling out: **scaffold compatibility.** Meta states explicitly that Muse Glimmer works with OpenClaw and other agentic orchestration patterns. If your stack already uses an agent orchestration framework (OpenClaw being the example Meta names), Muse Glimmer drops in as a plug-in without rewriting the scaffold.\n\n## 4. The \"open + closed\" double track behind this move\n\nPut Muse Glimmer (open-source 30B local agent) next to Muse Spark 1.2 (closed, released days earlier) and Meta's strategy becomes clear:\n\n- **Muse Spark \u002F Spark 1.2** (closed): frontier capability, enterprise API, agent platform.\n- **Muse Glimmer** (open-source 30B): local, edge, personal agent, research community.\n\nThis is the concrete productization of Meta's long-standing open AI strategy: **letting developers, researchers, and even enthusiasts run a 24\u002F7 personal AI assistant on their own machines.**\n\nThe partner roster also signals Meta's ambition: Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, SGLang, Together AI, Fireworks AI, OpenRouter — **full coverage from local to cloud.** On the hardware side, AMD, Arm, Dell, Intel, and NVIDIA are all in.\n\n## 5. So what?\n\nMuse Glimmer is not \"a bigger Llama.\" It is Meta's concentrated bet on the path of \"AI actually running on personal devices.\" With 30B + 4-bit quantization + speculative decoding as the formula, if the open-source community ships solid integrations for llama.cpp \u002F MLX \u002F ExecuTorch, **calendar management, file organization, code assistants, personal RAG retrieval** could be the first set of consumer-grade local agents that truly \"work offline, run fast, and run affordably.\"\n\nThe lesson for the Chinese open-model ecosystem is direct: **\"good-enough + runnable\" beats \"bigger\" in commercial value.** Meta's 30B-quantization-speculative-decoding-distillation stack draws a clear ceiling: **for 24\u002F7 local agent experience, 30B may be the sweet spot at this engineering node.** The next variable worth watching is whether the generation after DFlash can push that sweet spot even further down.\n\n---\n\n**Sources:**\n- Meta AI Research blog: https:\u002F\u002Fresearch.meta.ai\u002Fblog\u002Fintroducing-muse-glimmer-open-agentic-model\n- CNBC: https:\u002F\u002Fwww.cnbc.com\u002F2026\u002F08\u002F10\u002Fmeta-muse-glimmer-open-weight-ai.html","muse-glimmer-30b-open-agentic-local","2026-08-10T00:00:00Z","2026-08-10T14:13:47.260843Z","2026-08-10T14:13:47.260853Z",true,"agent",72]