🔶 Muse Glimmer：面向常驻本地 Agent 工作流优化的 300 亿参数模型
英文标题：🔶 Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows
Meta Superintelligence Labs 发布并以宽松的 Apache 2.0 许可证开源 Muse Glimmer 的模型权重。这是一款 300 亿参数模型，针对常驻运行的本地 Agent 工作流优化，可在配备单张消费级 GPU 的 Mac 或 PC 上运行。该模型通过蒸馏、量化和推测解码，在本地硬件上兼顾 Agent 能力、内存占用与响应速度，并支持文本与图像输入、多工具调用及超过 100 种语言。Meta 表示，开发者现可从 Hugging Face 下载模型，并将陆续通过 llama.cpp、MLX、ExecuTorch、Ollama 等生态工具进行部署和集成。
13.7 KB
13.7 KB
4.9 KB
4.9 KB
中文
今天，我们推出 Meta Superintelligence Labs 的下一款模型 Muse Glimmer，并以宽松的 Apache 2.0 许可证开源其模型权重。
Muse Glimmer 是一款 300 亿参数模型，针对常驻运行的本地 Agent 工作流进行了优化。它的体量足以在配备单张消费级 GPU 的 Mac 或 PC 上运行，可支持从本地 Agent、函数调用到本地编程和 LLM-as-a-judge 评估等多类场景。与同等规模类别中的领先模型相比，Muse Glimmer 在关键 Agent 用例和基准测试中表现强劲。
基础模型已在推理、代码生成和工具使用等领域展现出卓越能力——但大多数部署仍依赖云基础设施和网络连接。在本地运行模型，意味着无论是否联网，都可以随时随地使用 AI。这种模式正变得愈发可行：开源社区已经证明，经过有效训练的较小模型能够在特定任务上接近前沿模型的性能。Muse Glimmer 正是为这些本地使用场景而优化。
延续我们长期分享基础 AI 研究的传统，今天我们在 Hugging Face 发布 Muse Glimmer 的开放权重，同时提供开发者文档，帮助你开始构建和运行自己的 Agent。Muse Glimmer 旨在与开发者已在使用的工具协同工作。未来几天内，针对 llama.cpp、MLX 和 ExecuTorch 的优化集成将陆续上线，让你能够在几分钟内完成从下载到运行 Agent 的过程。
我们如何训练 Muse Glimmer
一个能够管理日程、起草消息、整理文件并学习你的工作方式的 Agent，需要深入访问个人上下文。它还需要多项能力协同发挥作用：长时程执行、精确工具调用、多模态理解、长上下文记忆和指令遵循。
我们将 Muse Glimmer 设计为在能力与本地硬件的内存和计算限制之间取得平衡。这需要紧凑的架构、可将 Agent 推理能力从更大型教师模型迁移而来的新型蒸馏方案，以及包括量化在内的推理优化，以满足延迟预期。我们通过以下阶段实现这一目标：
- 预训练。 我们利用 logit 蒸馏，在 Muse Spark 的输出上训练 Muse Glimmer，并采用与教师模型相似的数据混合方式。
- 中期训练。 我们使用包含更丰富推理轨迹的长上下文、Agent 密集型数据，以及自然数据，对模型进行训练。
- 后训练。 我们将监督微调与策略内蒸馏及强化学习相结合，覆盖通用、推理、编程和 Agent 领域。
Muse Glimmer 按照 Meta 高级 AI 扩展框架所规定的标准进行评估，并在所有相关类别中接受了开放权重发布评估。
为 Agent 打造：Muse Glimmer 能做什么
构建有效的 Agent，需要多项关键能力协同配合，以实现用户目标。Muse Glimmer 针对以下各项能力进行了训练和评估：
- 端到端 Agent 任务完成能力。 Muse Glimmer 在 DeepSearch QA、MCP-Atlas、𝛕-Bench 和 SWE-Bench 等完整任务基准中取得了较高成功率。这些基准衡量模型在脚手架环境中工作、编写和调试代码，以及从头到尾解决多轮请求的能力。
- 可靠的工具使用。 该模型能够处理广泛的函数调用，并在扩展工作流中依据精确的 schema 调用工具。
- 多步骤推理。 Muse Glimmer 可在长时程中串联推理，在复杂、延续性较强的工作流中保持连贯计划。
- 故障恢复。 当工具调用失败或返回意外结果时，模型经过训练能够诊断错误并重试，而非停止运行。
- 多模态输入与推理。 通过专用感知编码器，该模型可接受交错的文本和图像输入。这使 Agent 能够结合对话理解截图、图表和文档。
- 脚手架兼容性。 Muse Glimmer 可适配 OpenClaw 及其他 Agent 编排模式。
- 可控推理强度。 Muse Glimmer 支持不同级别的推理强度，以在质量和速度之间选择适当平衡。
- 多语言能力。 Muse Glimmer 基于超过 100 种语言的数据进行训练。
性能
我们在广泛的基准测试中评估了 Muse Glimmer，以衡量实现有效自主 Agent 行为所需的多样化能力。与 Gemma4-31B 和 Qwen3.6-27B 相比，Muse Glimmer 在多项广泛使用的 LLM 基准上，于其规模类别中表现强劲。
有关评估的更多详细信息，请参阅我们的报告。
针对本地部署优化
只有足够快、能让人感到响应及时的本地 Agent 才真正实用。一个需要数分钟才能回复或规划下一步的 Agent，会打断真实工作的节奏。我们采用了两项优化，使 Muse Glimmer 能够在不牺牲质量的前提下，以实用速度运行在消费级硬件上。
让模型适配你的设备
在全精度下，一个 300 亿参数模型需要超过 55 GB 内存——远超任何消费级 GPU 的容量。我们采用量化技术将模型权重压缩至约 4 位精度，将语言模型缩小到不足 20 GB。这为模型工作内存（即其“KV cache”）、用于图像理解的感知编码器，以及推测解码草稿模型同时运行于 24 GB 或 32 GB 内存范围内留出了充足空间。我们验证了，这种压缩在 Agent 任务上仅带来极小甚至没有性能下降。
通过推测解码加快生成
语言模型通常一次生成一个 token，在进行长推理链或多步骤工具调用时，这种方式可能显得缓慢。Muse Glimmer 随附基于 DFlash 的轻量级“草稿模型”——这是一个小型配套网络，可一次提出整块 token。主模型随后并行验证这些提议，接受正确的 token，并修正错误 token。这项技术使 Muse Glimmer 能够比标准的逐 token 生成显著更快地生成文本，同时产出相同质量的结果。我们还在此次发布中提供量化版草稿模型，以带来更小的额外内存开销。
最终结果：
我们在 MacBook M4-Max、M5-Max 以及 RTX-5090 上，测量了 K-Quant-17GB 模型搭配量化 DFlash 草稿模型时的速度。该模型速度足以支持流畅对话和实时 Agent 交互，且全部运行在你的设备上。
立即开始使用 Muse Glimmer
Muse Glimmer 现已可用，你可以在 Hugging Face 下载模型权重。未来几天内，你可以通过 Ollama、LM Studio 和 Unsloth 等合作伙伴在本地运行它；通过包括 llama.cpp、ExecuTorch 和 MLX 在内的边缘框架部署它；通过 vLLM 和 SGLang 大规模提供服务；或通过 Together AI、Fireworks AI 和 OpenRouter 等合作伙伴快速开始。你还可以利用 PyTorch 的 TorchTitan 训练功能进一步微调模型，使其适配自己的使用场景。
我们还正与 AMD、Arm、Dell、Intel 和 NVIDIA 等合作伙伴协作，优化模型在各类设备上的性能。此外，我们发布了文档，为开发者提供开始使用 Muse Glimmer 并以负责任方式进行构建所需的资源。其中包括有关设置自定义脚手架的指导，以便开发者从第一天起就能更轻松地构建和部署个人 Agent。你可以在 Meta AI Developer Center 了解更多信息，并获取构建所需资源。
这项工作建立在 Meta 长期开展开放 AI 研究的基础上，将这一传统延伸至 Agentic AI，并让开发者获得本地 Agent 能力。与往常一样，我们欢迎社区反馈，并期待看到开发者利用这款开放权重模型构建出什么。
English / 原文
Today, we're introducing Muse Glimmer, the next model from Meta Superintelligence Labs, and open sourcing the model weights under a permissive Apache 2.0 license.
Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows. It’s small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category.
Foundation models have achieved remarkable capabilities across reasoning, code generation, and tool use — yet most deployments still depend on cloud infrastructure and network access. Running models locally enables you to use AI anywhere, anytime, with or without an internet connection. This is increasingly viable: the open source community has shown that smaller models, when trained effectively, can approach frontier-level performance on targeted tasks. Muse Glimmer is optimized for these local use cases.
Keeping with our long tradition of sharing fundamental AI research, we're releasing Muse Glimmer open weights today on Hugging Face, along with developer documentation to help you start building and running your own agents. Muse Glimmer is built to work with the tools developers already use. Optimized integrations on llama.cpp, MLX, and ExecuTorch will land in the coming days, so you can go from download to working agent in minutes.
How We Trained Muse Glimmer
An agent that manages your schedule, drafts your messages, organizes your files, and learns how you work needs deep access to personal context. It also needs several capabilities working in concert: long-horizon execution, precise tool calling, multimodal understanding, long-context memory, and instruction following.
We designed Muse Glimmer to balance capability against the memory and compute constraints of local hardware. This required a compact architecture, a novel distillation recipe that transfers agentic reasoning from a much larger teacher model, and inference optimizations — including quantization — to meet latency expectations. We achieved this in the following phases:
- Pre-Training. We trained Muse Glimmer on Muse Spark's outputs using logit distillation, leveraging a similar data mix as the teacher.
- Mid-Training. We trained the model on longer-context, more agent-heavy data with richer reasoning traces, alongside organic data.
- Post-Training. We combined supervised fine-tuning with a mix of on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains.
Muse Glimmer was evaluated under the standards set out in Meta's Advanced AI Scaling Framework and assessed for open-weight release across all relevant categories.
Built for Agents: What Muse Glimmer Can Do
Building effective agents requires key capabilities working together to achieve the user’s goals. Muse Glimmer is trained and evaluated across each of the following:
- End-to-end Agentic Task Completion. Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure its ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish.
- Reliable Tool Use. The model handles a wide range of function calls, invoking tools with precise schemas throughout extended workflows.
- Multi-Step Reasoning. Muse Glimmer chains reasoning over long horizons, sustaining coherent plans across complex, extended workflows.
- Failure Recovery. When a tool call fails or returns an unexpected result, the model is trained to diagnose the error and retry rather than halt.
- Multimodal Input and Reasoning. Through a dedicated perception encoder, the model accepts interleaved text and images. This enables agents to interpret screenshots, charts, and documents alongside conversation.
- Scaffold Compatibility. Muse Glimmer works across OpenClaw and other agentic orchestration patterns.
- Controllable Effort. Muse Glimmer supports different reasoning strengths to select the right balance between quality and speed.
- Multilingual. Muse Glimmer is trained on data from more than 100 languages.
Performance
We evaluated Muse Glimmer across a broad range of benchmarks to assess the diverse capabilities required for effective autonomous agent behavior. Compared with Gemma4-31B and Qwen3.6-27B, Muse Glimmer performs strongly for its size class on several widely used LLM benchmarks.
For more detail about our evaluations, see our report.
Optimized for Local Deployments
A local agent is truly useful if it's fast enough to feel responsive. An agent that takes minutes to reply or plan its next step breaks the flow of real work. We applied two optimizations to make Muse Glimmer run at practical speeds on consumer hardware without sacrificing quality.
Fitting the Model on Your Device.
At full precision, a 30-billion parameter model would require over 55 GB of memory — far more than any consumer GPU offers. We use quantization techniques to compress the model's weights to approximately 4-bit precision, shrinking the language model to under 20 GB. This leaves enough headroom for the model's working memory (its "KV cache"), the perception encoder for image understanding, and the speculative decoding drafter to run simultaneously within a 24 GB or 32 GB envelope. We validated that this compression introduces minimal to no degradation on agentic tasks.
Faster Generation Through Speculative Decoding.
Language models normally generate text one token at a time, which can feel slow during long reasoning chains or multi-step tool calls. Muse Glimmer ships with a lightweight "drafter" model based on DFlash — a small companion network that proposes entire blocks of tokens at once. The main model then verifies these proposals in parallel, accepting correct tokens and correcting wrong ones. This technique lets Muse Glimmer generate text significantly faster than standard token-by-token generation while producing identical output quality. We provide quantized drafter versions to incur a smaller memory overhead in the release.
The Result:
We measure the speed of our K-Quant-17GB model alongside the quantized DFlash drafter on MacBook M4-Max, M5-Max and on a RTX-5090. The model is fast enough for fluid conversation and real-time agent interaction, all running entirely on your device.
Get Started With Muse Glimmer Today
Muse Glimmer is available now, and you can download the weights on Hugging Face. In the coming days, run it locally through partners like Ollama, LM Studio, and Unsloth, deploy it with edge frameworks including llama.cpp, ExecuTorch, and MLX, serve it at scale with vLLM and SGLang, or get started quickly through partners like Together AI, Fireworks AI, and OpenRouter. You can even customize it for your use case by leveraging PyTorch’s TorchTitan training feature to tune the model further.
We're also working with our partners including AMD, Arm, Dell, Intel, and NVIDIA to optimize performance across devices. In addition, we’re releasing documentation so developers have the resources they need to get started and build responsibly with Muse Glimmer. This includes guidance on setting up custom scaffolds, so it's even easier to start building and deploying personal agents on day one. You can learn more and find resources to build on Meta's AI Developer Center.
This work builds on Meta's long track record of open AI research, extending it into agentic AI and giving developers access to local agentic capabilities. As always, we welcome feedback from the community and can’t wait to see what developers build with this open weights model.