Meta 发布 30B 开源模型 Muse Glimmer，可在单张 NVIDIA GPU 上跑本地智能体
Run Local Agentic AI Workflows with Meta’s Muse Glimmer on NVIDIA | NVIDIA Technical Blog
Meta 发布 Muse Glimmer，一个 30B 开源权重稠密模型，上下文窗口 120K+，面向本地 AI 智能体工作负载。该模型采用稠密架构，每个 token 激活全部参数，在 NVIDIA Blackwell Ultra 上以 BF16/NVF4 精度实现单 GPU 超过 20K tokens/s，可完整放入单卡显存，无需模型分片、CPU offload 或外部端点，覆盖 GeForce RTX 5090、DGX Spark、DGX Station 和 Jetson 平台。开发者可用 NeMo AutoModel 做全量 SFT 和 LoRA 微调，权重已上线 Hugging Face，并提供 SGLang、vLLM 推理方案和可下载的 NIM 容器。
推荐理由
精选AI 评分 71
Meta 以 30B 稠密模型重返开源，配合单卡 20K tokens/s 的本地推理数据，可据此判断端侧智能体的可行边界。
主题
正文
当前显示原文，中文全文译文尚未就绪。
Technical Blog
sessionStorage.removeItem('facets.search.text'); sessionStorage.removeItem('facets.search.text-copy'); sessionStorage.removeItem('facets.search.filters');
Aug 10, 2026
Muse Glimmer is a 30B open-weight dense model with a 120K+ context window built for local AI agentic work. The dense architecture activates every parameter for each token, delivering reliable instruction following, long-context coherence, predictable latency, and fewer failure modes for multi-step agentic workloads. NVIDIA Tensor Core architecture accelerates Muse Glimmer to over 20K tokens per second on a single GPU, enabling real-time agentic inference fully on device at full context length. The model fits within the VRAM of a single NVIDIA GPU, eliminating the need for model sharding, CPU offloading, or external endpoints across GeForce RTX 5090, DGX Spark, DGX Station, and Jetson platforms. Developers can post-train Muse Glimmer using NeMo AutoModel for full SFT and LoRA fine-tuning with native Hugging Face checkpoint support. Next Steps Download the Muse Glimmer weights from HuggingFace and deploy using SGLang or vLLM inference recipes. Pull the downloadable NIM for a production-ready container on any NVIDIA GPU-accelerated platform. Try the hosted endpoint on build.nvidia.com for instant access.
Powered by NVIDIA Nemotron. AI-generated content may summarize information incompletely. Verify important information. Learn more
Meta returns to the open source ecosystem with the release of Muse Glimmer, a 30B open-weight dense model with a 120K+ context window built for local AI agentic work.
Optimized to run across a range of NVIDIA edge, desktop, and workstation AI platforms, Muse Glimmer delivers 20K tokens/sec on a single GPU, enabling always-on agents to process data locally and execute complex, multi-step workflows.
Most LLMs are optimized for chat , prioritizing single-turn interactions and fast time to first token—but agentic workloads demand a different approach. An agent scaffolding a software project, revising documentation, or managing a knowledge base may execute several sequential tool calls in a single session, while requiring a level of reliability, consistency, long-context coherence, and sustained throughput that chat-first models aren’t built for.
Muse Glimmer uses a dense architecture that activates every parameter for each token it processes, with no routing, expert selection, or variance across token pathways. As a result, it excels at agentic workloads that demand reliable instruction following, long-context coherence, predictable latency, and fewer failure modes.
Figure 1. Overview of the Muse Glimmer dense model architecture compared to MoE architecture
Privacy by design across local hardware
Agentic workflows involving personal files, communications, credentials, and proprietary documents require inference that never leaves the machine. Muse Glimmer hits an optimal balance. It’s large enough for complex multi-step reasoning, but small enough to fit within the VRAM of a single NVIDIA GPU, with no need for model sharding, CPU offloading, or using external endpoints.
NVIDIA Tensor Core architecture accelerates exactly this compute pattern, enabling real-time agentic inference fully on device at full context length.
On NVIDIA Blackwell Ultra, Muse Glimmer delivers over 20K tokens/sec/GPU at BF16/NVF4 precision, with the throughput-interactivity curve showing the 30B dense architecture sustaining high concurrency without the routing overhead of MoE models.
A single Blackwell Ultra handles the full model in VRAM with headroom for large KV cache buffers, making it well-suited for high throughput and low latency that developers need to run always-on agents entirely on local infrastructure.
Figure 2. Muse Glimmer performance on NVIDIA Blackwell Ultra throughput at BF16 precision
Building and fine-tuning agentic use cases
Run NVIDIA NemoClaw in a secure OpenShell environment to create long-running personal assistants powered for tasks like code generation, personal assistant, autonomous support, and more.
Figure 3. Muse Glimmer running locally with the NemoClaw agent harness in a governed sandbox, served by vLLM on DGX Spark
Developers can further post-train the model using the NVIDIA NeMo AutoModel with high-throughput efficiency, which is a fine-tuning library for native Hugging Face checkpoint support with no model conversion requirements.
It enables full SFT and LoRA fine-tuning out of the box, optimized for rapid experimentation on NVIDIA GPUs, including DGX Spark. Developers can also perform reinforcement learning with NeMo RL, with sample recipes and reference accuracy validation curves .
NVIDIA supports multiple inference stacks to meet a variety of developer needs.
SGLang and vLLM provide open-source inference recipes for developers who require deeper control over performance on the NVIDIA accelerated platform.
It’s also available as a downloadable NVIDIA NIM , a prebuilt, optimized inference container that auto-selects runtime configuration and serving setup, so teams can focus on building and scaling agents.