Meta's Muse Glimmer: 30B open-weight model for local AI agents

Meta has released Muse Glimmer, a 30B open-weight dense model with a 120K+ context window, optimized for local AI agentic workflows on NVIDIA hardware. The model delivers 20,000 tokens per second on a single GPU and is designed to enable always-on agents to process data and execute complex multi-step workflows entirely on-device.

Muse Glimmer's architecture reflects a fundamental difference from chat-focused LLMs. Most large language models are optimized for chat, prioritizing single-turn interactions and fast time to first token. Agentic workloads, by contrast, demand agents that can execute several sequential tool calls in a single session—such as scaffolding a software project, revising documentation, or managing a knowledge base—while maintaining reliability, consistency, and sustained throughput. Muse Glimmer uses a dense architecture that activates every parameter for each token it processes, with no routing or expert selection. This design eliminates variance across token pathways and reduces failure modes, enabling reliable instruction following, long-context coherence, and predictable latency.

The model is built for privacy by design. Because agentic workflows often involve personal files, communications, credentials, and proprietary documents, inference must never leave the machine. Muse Glimmer is large enough for complex multi-step reasoning but small enough to fit within the VRAM of a single NVIDIA GPU, with no model sharding or CPU offloading required. On NVIDIA Blackwell Ultra, the model delivers over 20,000 tokens per second at BF16/NVF4 precision, with a single Blackwell Ultra handling the full model in VRAM with headroom for large key-value cache buffers. NVIDIA GeForce RTX 5090 pairs 32 GB of VRAM with fifth-generation Tensor Cores, bringing Muse Glimmer to local developer devices and eliminating per-token inference costs. NVIDIA DGX Spark brings workstation-class performance for enterprise agentic pipelines into a compact system, while NVIDIA DGX Station brings rack-scale Blackwell Ultra compute to on-premises enterprise environments for teams under air-gap mandates or strict compliance frameworks. NVIDIA Jetson extends local Muse Glimmer inference to edge devices, enabling robotics, industrial automation, and embedded systems.

Developers can build and fine-tune agentic use cases using NVIDIA NemoClaw in a secure OpenShell environment to create long-running personal assistants for code generation, autonomous support, and other tasks. The NVIDIA NeMo AutoModel enables full supervised fine-tuning and LoRA fine-tuning out of the box, optimized for rapid experimentation on NVIDIA GPUs. Reinforcement learning is supported via NeMo RL, with sample recipes and reference accuracy validation curves. Deployment paths include open-source inference recipes like SGLang and vLLM for developers requiring deeper control over performance, as well as a downloadable NVIDIA NIM—a prebuilt, optimized inference container that auto-selects runtime configuration and serving setup. Developers can download Muse Glimmer weights from HuggingFace, deploy using the inference recipes, or pull the NIM for production-ready deployment on any NVIDIA GPU-accelerated platform. A hosted endpoint is available on build.nvidia.com, and edge deployments on Jetson are supported through Jetson AI Lab.