[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-muse-glimmer-executorch-local-toolchain":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},"1048efcf-5071-450d-b7e2-08f2986c4139","Muse Glimmer 30B 本地工具链落地:ExecuTorch 官方支持,17GB 量化权重几条命令起 OpenAI 兼容服务","Meta 为 Muse Glimmer 30B 兑现了发布时承诺的本地部署工具链:pytorch\u002Fexecutorch 主分支新增官方示例,量化 GGUF 权重可直接导出或下载预构建 PTE,支持 CUDA 与 Apple silicon,自带 OpenAI 兼容 serving 与 atem 工具调用解析。30B 本地智能体模型从「能下载」走到了「能投产」。","**从发布到能跑:Muse Glimmer 的部署故事进入第二章**\n\nMeta 把 300 亿参数的开放权重模型 Muse Glimmer(Apache 2.0 许可证)开源时,定位就很明确:为本地运行的智能体工作流优化,配备单个消费级 GPU 的 Mac 或 PC 就能跑。但当时承诺的 llama.cpp、MLX、ExecuTorch 集成还写着「未来几天」。现在承诺兑现了——pytorch\u002Fexecutorch 主分支已经出现 muse-glimmer 官方示例目录,Meta 的开发者文档(dev.meta.ai)也同步上线了部署指南,从模型权重到本地服务的整条链路打通。\n\n**ExecuTorch 官方集成:GGUF 直通导出**\n\n这次落地的核心是 ExecuTorch 的官方支持。导出管线直接消费量化后的 GGUF checkpoint,产出 ExecuTorch 的 .pte 程序,支持 target-only 和 DFlash 投机解码两种导出模式,可选带上视觉输入。后端方面支持 CUDA(Linux 或 Windows)和 MLX(macOS Apple silicon),CPU 导出则明确不支持。\n\n官方推荐的起点是 Hugging Face 上的 meta-models\u002FMuse-Glimmer-30B-GGUF 仓库,首选 checkpoint 是 Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf——一个 17GB 的 KQuant 量化版本;另有动态 K-quant 的 Q4_K_XL 版本、可选的视觉投影 mmproj 文件,以及用于投机解码的 DFlash draft checkpoint。\n\n**预构建 PTE:连导出这步都省了**\n\n更值得注意的一步:Meta 直接在 meta-models\u002FMuse-Glimmer-30B-ExecuTorch-PTE 仓库发布了预构建导出。每个子目录就是一个独立导出,按量化方式、上下文长度、模态、解码模式(solo 或 dflash)和目标硬件命名,比如 muse_glimmer_k_quant_17G_128K_text_solo_metal——17GB 量化、128K(131072)上下文、纯文本、单模型解码、Apple metal 后端。不想自己走导出流程的开发者,挑一个匹配硬件的子目录下载就能跑。\n\n**Serving 与智能体接入**\n\n运行时侧,官方示例自带 OpenAI 兼容 serving,启动后直接 curl \u002Fv1\u002Fchat\u002Fcompletions 做冒烟测试;工具调用通过 atem 解析器把模型的 ATEM 输出转换成 OpenAI 风格的 tool_calls。README 里甚至给了 pi 编码代理的接入配置:contextWindow 设为 131072,reasoning 输出通道(to=self)映射为 reasoning_content,配合 --tools read,bash,edit,write 就能当一个本地智能体后端用。在 Meta 的部署文档里,ExecuTorch 也只是四条路径之一——vLLM、SGLang、llama.cpp 同样在列,生产侧和端侧各有分工。\n\n**评论:开源模型的分水岭在工具链**\n\n模型权重只是入场券。Muse Glimmer 这次补齐的工具链说明,开放权重模型真正的竞争点在「从下载到投产要几步」:GGUF 直接降级到 ExecuTorch 后端,意味着 llama.cpp 系的量化生态和 PyTorch 端侧运行时不再各玩各的;预构建 PTE 则把最后一步导出成本也砍掉了。对开发者来说,现在在一台 Apple silicon Mac 或单张 NVIDIA GPU 上,一个 17GB 的 30B 智能体模型,从下载到起一个 OpenAI 兼容服务,已经是几条命令的事。下一个值得观察的问题:这条 GGUF 到 PTE 的直通管线,会不会成为后续大模型端侧部署的标准动作。\n\n参考:[pytorch\u002Fexecutorch muse-glimmer 示例](https:\u002F\u002Fgithub.com\u002Fpytorch\u002Fexecutorch\u002Ftree\u002Fmain\u002Fexamples\u002Fmodels\u002Fmuse-glimmer)","https:\u002F\u002Fgithub.com\u002Fpytorch\u002Fexecutorch\u002Ftree\u002Fmain\u002Fexamples\u002Fmodels\u002Fmuse-glimmer","998df6db-96e6-4b8e-8be1-cfa00a6cd177",[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},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"7e24e4d4-131d-4f01-8552-f7f8fbb08778","en","Muse Glimmer 30B Local Toolchain Lands: Official ExecuTorch Support, a 17GB Quantized Checkpoint, and an OpenAI-Compatible Server in a Few Commands","Meta has delivered the local deployment toolchain it promised at launch for Muse Glimmer 30B: the pytorch\u002Fexecutorch main branch now ships an official muse-glimmer example, quantized GGUF checkpoints export directly to PTE or arrive as prebuilt artifacts, CUDA and Apple silicon are both supported, and OpenAI-compatible serving with the atem tool-call parser is included. A 30B agentic model has moved from \"downloadable\" to \"deployable.\"","**From Release to Runnable: Muse Glimmer's Deployment Story Enters Chapter Two**\n\nWhen Meta open-sourced Muse Glimmer, its 30-billion-parameter open-weights model under the Apache 2.0 license, the positioning was explicit: optimized for locally running agentic workflows, able to run on a Mac or PC equipped with a single consumer-grade GPU. But the llama.cpp, MLX, and ExecuTorch integrations promised at launch were still tagged \"in the next few days.\" That promise has now been fulfilled — the pytorch\u002Fexecutorch main branch carries an official muse-glimmer example directory, Meta's developer docs (dev.meta.ai) launched the deployment guide alongside it, and the full pipeline from model weights to a local server is connected.\n\n**Official ExecuTorch Integration: GGUF Pass-Through Export**\n\nThe core of this rollout is official ExecuTorch support. The export pipeline consumes quantized GGUF checkpoints directly and produces ExecuTorch .pte programs. It supports target-only and DFlash speculative-decoding exports, with optional vision input. On the backend side, CUDA (Linux or Windows) and MLX (macOS on Apple silicon) are supported, while CPU export is explicitly not supported.\n\nThe recommended starting point is the meta-models\u002FMuse-Glimmer-30B-GGUF repository on Hugging Face, with the Muse-Glimmer-30B-KQuant-17GB-Q4_K_M.gguf checkpoint as the recommended target — a 17GB KQuant quantization. There is also a dynamic K-quant Q4_K_XL variant, an optional vision projector (mmproj) file, and a DFlash draft checkpoint for speculative decoding.\n\n**Prebuilt PTE Artifacts: The Export Step Is Now Optional**\n\nThe more notable step: Meta published prebuilt exports in the meta-models\u002FMuse-Glimmer-30B-ExecuTorch-PTE repository. Each subdirectory is one export, named after its quantization, context length, modalities, decoding mode (solo or dflash), and target hardware — for example muse_glimmer_k_quant_17G_128K_text_solo_metal: 17GB quantization, 128K (131072) context, text-only, solo decoding, Apple metal backend. Developers who don't want to run the export pipeline themselves can pick the subdirectory that matches their hardware, download it, and go.\n\n**Serving and Agent Integration**\n\nOn the runtime side, the official example ships an OpenAI-compatible server — after launch, a curl against \u002Fv1\u002Fchat\u002Fcompletions serves as the smoke test. Tool calling works through the atem parser, which converts the model's ATEM output into OpenAI-style tool_calls. The README even includes a configuration for the pi coding agent: set contextWindow to 131072, map the reasoning channel (to=self) to reasoning_content, and with --tools read,bash,edit,write the model serves as a local agent backend. In Meta's deployment docs, ExecuTorch is only one of four paths — vLLM, SGLang, and llama.cpp are all listed too, splitting the work between production and on-device deployment.\n\n**Commentary: The Real Divide for Open-Weights Models Is the Toolchain**\n\nModel weights are just the entry ticket. The toolchain Muse Glimmer just completed shows that the real competitive axis for open-weights models is \"how many steps from download to production\": lowering GGUF directly to an ExecuTorch backend means the llama.cpp quantization ecosystem and the PyTorch on-device runtime are no longer playing separate games, and the prebuilt PTE artifacts eliminate the last export cost. For developers, on an Apple silicon Mac or a single NVIDIA GPU, taking a 17GB 30B agentic model from download to a running OpenAI-compatible server is now a matter of a few commands. The next question worth watching: will this GGUF-to-PTE pass-through pipeline become the standard move for on-device deployment of future large models?\n\nReference: [pytorch\u002Fexecutorch muse-glimmer example](https:\u002F\u002Fgithub.com\u002Fpytorch\u002Fexecutorch\u002Ftree\u002Fmain\u002Fexamples\u002Fmodels\u002Fmuse-glimmer)","muse-glimmer-executorch-local-toolchain","2026-08-15T21:00:00Z","2026-08-15T21:08:25.993733Z","2026-08-15T21:08:25.993742Z",true,"agent",36]