Meta Open-Sources Muse Glimmer: A 30B Agentic Model That Runs on One Consumer GPU

Meta releases Muse Glimmer today: 30B open-weights agentic model (Apache 2.0), distilled for tool use and failure recovery, fits on a single consumer GPU at 18-20 GB with 4-bit quantization.

Muse Glimmer is a 30-billion-parameter dense multimodal model distilled from Muse Spark 1.2 and specifically trained for agentic workloads: multi-step tool use, function calling, LLM-as-a-judge evaluations, and task failure recovery. Full weights are 55 GB; Unsloth's GGUF 4-bit build (already on HuggingFace as of today) cuts that to 18–20 GB, within reach of a single RTX 4090 or a Mac with 24 GB+ unified memory.

Native integrations are available today for llama.cpp, MLX, ExecuTorch, and Ollama (ollama pull muse-glimmer). Text and image input supported.

Why it matters: Until today, every open model that could handle real agentic workloads reliably (Kimi K3, DeepSeek V3.2, the Muse Spark series) required multi-GPU infra or cloud APIs. Muse Glimmer is the first credible open agentic model that fits on prosumer hardware — making local agent development and self-hosted production deployments genuinely viable.

Sources: Introducing Muse Glimmer — Meta AI Research · meta-models/Muse-Glimmer-30B — HuggingFace · unsloth/Muse-Glimmer-30B-GGUF — HuggingFace · Muse Glimmer on Phoronix