Meta releases Muse Glimmer, an open-weight 30B local agent model
Weights are available under Apache 2.0; the quantized model runs entirely on consumer hardware like a Mac or PCs with performant GPUs.
- Quantized to under 20GB and paired with a lightweight DFlash drafter model to keep token generation fast enough for real-time agent interaction on-device.
- Completes multi-step agentic tasks from a single natural-language prompt; the demo shows it discovering a local Home Assistant instance, writing an HTML/CSS/JS dashboard, and deploying a verification server.
- Claims strong performance on agentic benchmarks versus leading models in its size category.
- Weights are downloadable on Hugging Face (meta-models), with a technical blog and developer resources linked.
Developers can now run a 30B agentic model entirely on consumer hardware under a permissive license, enabling private, always-on agent workflows without cloud dependency.
Sources
Read this as text
Back to the AI news