Meta Launches Muse Glimmer for Local Agentic AI

Meta has introduced Muse Glimmer, an open-weight 30-billion-parameter dense model featuring a context window of over 120,000 tokens. Unlike mixture-of-experts architectures that route tokens to specific sub-networks, Muse Glimmer activates every parameter for every token. This dense design provides the predictable latency, long-context coherence, and reliable instruction-following required for multi-step agentic tasks.

The model is optimized to run locally across a spectrum of Nvidia hardware. On Nvidia Blackwell Ultra GPUs, Muse Glimmer achieves a throughput of over 20,000 tokens per second per GPU at BF16 and NVF4 precision. This performance allows a single GPU to hold the entire model in VRAM alongside a large key-value cache. For local developers, the model runs on the Nvidia GeForce RTX 5090, which features 32 GB of VRAM and fifth-generation Tensor Cores, eliminating per-token cloud costs and keeping proprietary data secure.

For enterprise environments, the model deploys on DGX Spark workstations and DGX Station systems featuring Blackwell Ultra compute, which are ideal for air-gapped setups. It also runs on Nvidia Jetson modules to power edge robotics and industrial automation. Developers can deploy the model using open-source engines like SGLang and vLLM, or package it instantly via downloadable Nvidia NIM containers.

To build agentic applications, developers can run the NemoClaw agent harness in a secure OpenShell sandbox. For customization, the Nvidia NeMo AutoModel library supports supervised fine-tuning and Low-Rank Adaptation directly on Hugging Face checkpoints without requiring model conversion. Reinforcement learning is also supported through NeMo RL. These tools allow practitioners to build highly customized, always-on local assistants for software automation and document management.