Meta Open-Sources MuseGlimmer: A 30B Local Agent Model Running on 24GB VRAM

大模型之路 · wechat · 2026-08-12

Meta's Superintelligence Lab has open-sourced MuseGlimmer (Apache 2.0), a 30B parameter multimodal model designed specifically as a resident local agent rather than a chatbot, emphasizing offline use and data privacy.

Architecture & Optimization

The model features a 27.9B text decoder and a 1.9B vision encoder, supporting 128k+ context. It employs a hybrid sliding window attention mechanism to balance long-context needs with compute costs. Using 4-bit quantization, weights are compressed to under 20GB, allowing it to run on 24GB VRAM GPUs (RTX 3090/4090) or 32GB Macs. It also includes a built-in draft model for speculative decoding to accelerate generation.

Ecosystem & Positioning

MuseGlimmer received day-0 support from major toolchains like llama.cpp, Ollama, and HuggingFace. Meta aims to bridge the gap between capable but heavy models and lightweight but weak ones. Compared to peers like Qwen3.6-27B, its edge lies in being natively optimized for on-device agentic tasks. This move signals a shift in the LLM landscape towards local intelligence and device-side compute.

Related event: Meta Open-Sources 30B Local Agent Model Muse Glimmer(2 posts)→

Original post →

More from Models

Models channel →