Meta Releases Muse Glimmer 30B Dense Model for Local Agentic Workflows on NVIDIA Hardware

Meta Releases Muse Glimmer 30B Dense Model for Local Agentic Workflows on NVIDIA Hardware Meta has released Muse Glimmer, an open-weight 30B dense model featuring a 120K+ context window tailored for local AI agent workflows across NVIDIA desktop, edge, and enterprise platforms. Muse Glimmer uses a dense architecture with a 120K+ context window that fits within the VRAM of a single NVIDIA GPU without sharding or offloading. On NVIDIA Blackwell Ultra, the model reaches over 20,000 tokens per second per GPU at BF16/NVF4 precision. Supported hardware includes the GeForce RTX 5090, DGX Spark, DGX Station, and Jetson edge systems. Developers can fine-tune with NVIDIA NeMo AutoModel and NeMo RL, and deploy using SGLang, vLLM, or NVIDIA NIM containers. Dense Architecture for Agentic Demands Unlike chat-optimized models prioritized for quick single-turn responses, Muse Glimmer activates every parameter for each token with no mixture-of-experts routing or pathway variance. This dense design supports long-context coherence, predictable latency, sustained throughput, and multi-step tool calls needed for persistent agent operations. Hardware Support and Local Privacy The 30B model can run locally on single GPUs to keep sensitive data on-premises. Hardware targets include the GeForce RTX 5090 with 32 GB of VRAM, the DGX Spark workstation, the rack-scale DGX Station for air-gapped enterprise environments, and Jetson platforms for network-isolated robotics and edge automation. On Blackwell Ultra, it delivers over 20K tokens per second per GPU at BF16/NVF4 precision while retaining memory headroom for KV cache buffers. Fine-Tuning and Deployment Ecosystem Agent applications can be run with NVIDIA NemoClaw inside an OpenShell environment. Fine-tuning is supported through NVIDIA NeMo AutoModel for SFT and LoRA directly on Hugging Face checkpoints without format conversion, alongside reinforcement learning via NeMo RL. For deployment, developers can use open-source stacks like SGLang and vLLM, download prebuilt NVIDIA NIM inference containers, or access hosted endpoints.

Meta发布30B开源模型Muse Glimmer，支持在NVIDIA硬件上本地运行Agent工作流 Meta推出专为本地AI Agent设计的30B开源稠密模型Muse Glimmer，具备120K+上下文窗口。该模型适配NVIDIA边缘、桌面和工作站平台，在单个GPU上可达20K tokens/sec吞吐量，以保障本地数据隐私和多步长程推理。 Muse Glimmer采用30B全激活稠密架构与120K+上下文窗口，专为长流程Agent任务设计。 在NVIDIA Blackwell Ultra单GPU上（BF16/NVF4精度），吞吐量超过20K tokens/sec。 模型可在单个NVIDIA GPU显存中完全容纳，适配RTX 5090、DGX Spark、DGX Station和Jetson硬件。 提供NVIDIA NeMo系列微调工具支持，并可通过vLLM、SGLang或NVIDIA NIM容器进行部署。 专为长运行Agent打造的稠密架构 与优先优化单轮对话和首字延迟的聊天模型不同，Agent需要执行长上下文连续工具调用。Muse Glimmer采用为每个token激活全部参数的稠密架构，无需路由或专家选择，减少了token路径差异，能够提供更稳定的指令遵循能力、连贯的长文本推理和可预测的延迟。 全本地硬件适配与隐私保护 为满足敏感代码和文件在设备端处理的需求，Muse Glimmer无需模型分片或CPU卸载即可装入单张NVIDIA GPU显存。该模型支持从配备32GB显存的GeForce RTX 5090、DGX Spark工作站、机架级DGX Station，一直到网络隔离环境下的Jetson边缘设备。 性能表现、微调与部署生态 在NVIDIA Blackwell Ultra上以BF16/NVF4精度运行时，单卡即可提供超20K tokens/sec吞吐并留有充足的KV缓存空间。开发人员可通过NeMo AutoModel直接对Hugging Face权重执行SFT和LoRA微调，使用NeMo RL进行强化学习，并借助SGLang、vLLM开源方案或预构建的NVIDIA NIM容器快速部署上线。

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. ## Built for long-running agents, not just conversations  Most LLMs are [optimized for chat](https://developer.nvidia.com/blog/building-for-the-rising-complexity-of-agentic-systems-with-extreme-co-design/), 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. ## 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. - NVIDIA [GeForce RTX 5090](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/) pairs 32 GB of VRAM with fifth-generation Tensor Cores, bringing Muse Glimmer to local developer devices, keeping proprietary code on device, and eliminating per-token inference cost. - NVIDIA [DGX Spark](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) brings workstation-class performance and enterprise agentic pipelines into a compact system. NVIDIA NVLink provides high-speed access to memory, and NVIDIA NIM containers make local Muse Glimmer deployment a one-command operation. - NVIDIA [DGX Station](https://www.nvidia.com/en-us/products/workstations/dgx-station/) brings rack-scale Blackwell Ultra compute to on-prem enterprise environments for teams operating under air-gap mandates or compliance frameworks where cloud inference isn’t an option. - NVIDIA [Jetson](https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/) extends local Muse Glimmer inference to the edge, enabling robotics, industrial automation, and embedded systems, where network isolation is a hard requirement, and every inference decision must happen at the point of action. ## Optimized Muse Glimmer Performance on NVIDIA Blackwell Ultra  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. ## Building and fine-tuning agentic use cases  Run [NVIDIA NemoClaw](https://www.nvidia.com/en-us/ai/nemoclaw/) in a secure OpenShell environment to create long-running personal assistants powered for tasks like code generation, personal assistant, autonomous support, and more. Developers can further post-train the model using the NVIDIA [NeMo AutoModel](https://github.com/NVIDIA-NeMo/Automodel/blob/main/docs/model-coverage/vlm/muse/muse_glimmer.mdx) 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](https://github.com/NVIDIA-NeMo/RL/tree/muse-glimmer-support) and reference accuracy [validation curves](https://github.com/NVIDIA-NeMo/RL/blob/muse-glimmer-support/docs/guides/models/muse-glimmer.md). ## Flexible deployment paths for Muse Glimmer  NVIDIA supports multiple inference stacks to meet a variety of developer needs. [SGLang](https://docs.sglang.io/cookbook/autoregressive/Meta/MuseGlimmer) and [vLLM](https://recipes.vllm.ai/meta-models/Muse-Glimmer-30B) 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](https://catalog.ngc.nvidia.com/orgs/nim/meta/containers/muse-glimmer/-), a prebuilt, optimized inference container that auto-selects runtime configuration and serving setup, so teams can focus on building and scaling agents. ## Get started with Muse Glimmer and local AI agents   To get started, [download Muse Glimmer weights from Hugging Face](https://huggingface.co/nvidia/Muse-Glimmer-30B-NVFP4) and deploy using the inference recipes above, or pull the [downloadable NIM](https://catalog.ngc.nvidia.com/orgs/nim/meta/containers/muse-glimmer/) for a production-ready container on any NVIDIA GPU-accelerated platform. To call a hosted endpoint instantly, try it on [build.nvidia.com.](https://build.nvidia.com/meta/muse-glimmer-30b) For edge deployments on Jetson, explore the [Jetson AI Lab.](https://www.jetson-ai-lab.com/models/muse-glimmer-30b/?ncid=so-link-308924-vt48)