Bijan Bowen tests Meta's 30-billion-parameter Muse Glimmer model through a quantized local setup on a high-end consumer GPU. The Apache 2.0 release offers a 131,000-token context window and speculative decoding, and it responds quickly enough to make the return of a serious open-weight Meta model feel notable even before the benchmark results are considered.
The practical results are mixed. Muse Glimmer produces a usable browser-style operating system and eventually repairs a three-dimensional subway scene, but a C++ skateboard project freezes the machine and needs manual cleanup. Its computer-aided-design work is uneven, and a luxury watch page misses important visual details.
Image-to-site reconstruction and generated SVG assets are more convincing, especially given the model's speed and local footprint. Bowen's overall assessment is that the release is exciting because Meta is competing in open weights again, while the tested model itself still needs careful supervision on demanding tasks.
Watch the original on YouTube