I am running it on a single RTX 3090 (24GB VRAM).
Some folks on Reddit are having the same experience: https://www.reddit.com/r/LocalLLaMA/comments/1vkm42m/muse_gl...
It uses an order of magnitude less VRAM at longer contexts which is a huge advantage over Qwen 3.6 27B