Sam Witteveen · 2026-08-10 · notable
Sam Witteveen — 'Meta's Open Weight: Muse Glimmer 30B'
Sam Witteveen's new video walks through Muse Glimmer, the ~29.6B Apache-2.0 agentic model Meta released today. It takes text and images, holds 131,072+ tokens of context, and fits in 17-20 GB once quantized.
Sam Witteveen walks through Meta's Muse Glimmer, a 30B Apache-2.0 agentic model built to run on one consumer GPU.
What is it?
Muse Glimmer is the ~29.6B-parameter open-weight model Meta published on 2026-08-10, and this video is Sam Witteveen's same-day walkthrough of it. The model is Apache-2.0, takes text and images as input, and is aimed at always-on local agent work rather than cloud serving.
How does it work?
The architecture is a dense causal transformer with a separate ~1.8B perception encoder that turns each image into up to 4,096 visual tokens. Context runs to 131,072+ tokens, and a bundled DFlash drafter does speculative decoding for a 3.1x speedup on an RTX 5090 and 1.5x-1.8x on Apple silicon.
Why does it matter?
Quantized builds bring Muse Glimmer down to 17-20 GB, so an RTX 5090 or an M4/M5 Max Mac can run the whole agent loop with no API bill. Sam Witteveen's channel is where many local-LLM builders decide what to download first, so a same-day review shapes how fast the weights spread.
Who is it for?
local-LLM builders, agent developers
Try it
Watch the video, then pull `meta-models/Muse-Glimmer-30B` from Hugging Face