Meta Releases Muse Glimmer, 30-Billion-Parameter Open-Weights Model

Meta released Muse Glimmer, an open-weights language model with 30 billion parameters that runs on personal computers. Through 4-bit quantization that compressed each weight, Meta reduced the memory footprint to under 20 gigabytes, enabling execution on Macs with a single consumer-grade graphics card. The model uses speculative decoding, where a less advanced "drafter" model generates an initial response that is then verified and refined. Meta trained Muse Glimmer on data from the proprietary Muse Spark series, followed by two additional training passes to improve long-prompt capabilities and agent task performance. The model includes adjustable "reasoning strength" settings.