Meta has released Muse Glimmer, an open-weight 30 billion parameter (30B) dense language model that supports over a 120,000-token context window. The model is positioned for agentic workloads that run locally—long-running personal assistants, multi-step tool use, documentation revision, and knowledge base management.
Designed for agentic workloads rather than chat-first interactions
Many large language models are optimized for chat: single-turn interactions and fast time-to-first-token. Agentic tasks have different requirements: sustained throughput, long-context coherence, reliable instruction-following, predictable latency, and fewer failure modes during extended sessions. Muse Glimmer’s dense architecture activates all model parameters for each token (no routing or expert selection), which the developers say yields more consistent behavior for these use cases.
Performance claims and context
The announcement states that Muse Glimmer delivers 20K tokens/sec on a single GPU, enabling always-on agents to process data locally and execute complex multi-step workflows. Later in the same material, NVIDIA’s Blackwell Ultra numbers are reported as “over 20 tokens/sec/GPU” at BF16/NVF4 precision, with the 30B dense model able to sustain high concurrency without the routing overhead associated with Mixture of Experts (MoE) architectures.
The source document does not reconcile these two numerical statements; both are reported here as presented in the vendor materials.
Privacy and on-device inference
The model is described as balancing size and deployability: large enough for complex multi-step reasoning yet small enough to fit in the VRAM of a single NVIDIA GPU without model sharding, CPU offload, or external endpoints. NVIDIA’s Tensor Core architecture is highlighted as accelerating the compute pattern Muse Glimmer uses, enabling on-device, full-context-length inference in real time.
The release names several NVIDIA platforms and their intended roles:
- NVIDIA GeForce RTX 5090: 32 GB VRAM with fifth-generation Tensor Cores, aimed at local developer devices to keep proprietary code on-device and avoid per-token inference costs.
- NVIDIA DGX Spark: workstation-class performance for enterprise agentic pipelines, with NVLink for high-speed memory access and NVIDIA NIM containers to simplify local deployment.
- NVIDIA DGX Station: provides rack-scale Blackwell Ultra compute for on-premises, air-gapped, or compliance-sensitive environments.
- NVIDIA Jetson: extends local Muse Glimmer inference to edge, robotics, industrial automation, and embedded systems where network isolation is required.
Blackwell Ultra throughput
According to the materials, Muse Glimmer on NVIDIA Blackwell Ultra achieves over 20 tokens/sec per GPU at BF16/NVF4 precision. A single Blackwell Ultra can host the entire model in VRAM while leaving room for large key-value cache buffers, which the vendor says supports the low latency and high throughput required for always-on local agents.
Fine-tuning and building agentic applications
NVIDIA’s tooling is presented for developers who want to build and fine-tune agentic use cases:
- NVIDIA NeMoClaw: an agent harness that can run in a secure OpenShell environment to create long-running personal assistants and other autonomous support systems.
- NVIDIA NeMo AutoModel: a fine-tuning library with native Hugging Face checkpoint support, enabling SFT and LoRA out of the box and optimized for high-throughput experimentation on NVIDIA GPUs, including DGX Spark.
- NeMo RL: reinforcement learning tooling with sample recipes and reference validation curves for RL-based tuning.
The announcement includes an example of Muse Glimmer running locally with the NeMoClaw agent harness served by vLLM on DGX Spark in a governed sandbox.
Deployment options and getting started
NVIDIA offers several inference stacks to meet different developer needs. SGLang and vLLM provide open-source inference recipes for teams that want fine-grained control over performance. There is also a downloadable NVIDIA NIM (inference container) that auto-selects runtime configuration and serving setup to simplify deployment.
Muse Glimmer weights are available for download from HuggingFace, and developers are encouraged to deploy using the listed inference recipes, the downloadable NIM, or to try the model via build.nvidia.com on NVIDIA GPU-accelerated platforms.
Summary
Muse Glimmer is a 30B dense model intended for long-context, agentic workloads that can run locally on NVIDIA hardware. The vendor materials describe multiple deployment paths, performance figures, and tooling for fine-tuning and agent development. The announcement includes differing performance figures in separate sections (20K tokens/sec on a single GPU vs. "over 20 tokens/sec/GPU" on Blackwell Ultra); both are reported here as stated in the source materials.
