Meta Open-Sources MuseGlimmer: A 30B Local Agent Model

AGI Hunt · wechat · 2026-08-10

Meta has open-sourced MuseGlimmer, a 30B parameter agent model (Apache 2.0). Designed for local execution, it uses 4-bit quantization to compress memory requirements under 20GB, fitting consumer GPUs with 24GB VRAM. It employs a lightweight DFlash draft model for speculative decoding, significantly accelerating inference with minimal quality loss.

MuseGlimmer outperforms Qwen3.6-27B in tool calling and retrieval benchmarks but falls slightly behind in desktop and terminal operations. Specifically trained for agentic loops, the model features self-healing capabilities for error recovery. Weights are available on HuggingFace, with support for major inference frameworks rolling out.

Related event: Meta Releases Muse Glimmer 30B Open-Source Model for Local Agents(30 posts)→

Original post →

More from coding & agent

coding & agent channel →