Meta released Muse Glimmer on Monday, a 30-billion-parameter open-weight agentic model designed to run on consumer GPUs and laptops. The model is available for free download on Hugging Face under an Apache 2.0 license and optimized for always-on local workflows—schedule management, file organization, coding tasks, and agent-based tool calling without cloud dependency. Meta used quantization techniques to compress Glimmer to approximately 17 gigabytes, down from 55+ gigabytes in full precision, making it compatible with consumer GPUs containing 24–32GB of memory.
Glimmer was created through logit distillation from Meta's larger Muse Spark 1.2 model, transferred reasoning capabilities for agentic tasks, and added a lightweight speculative-decoding layer ('DFlash drafter') that predicts 16-token blocks. The drafter increased generation speed from 74.9 to 233.4 tokens/second on an RTX 5090, and from 23.7 to 37.8 on an M4 Max. Meta reports Glimmer scored 77.0 on IFBench (vs. Gemma4-31B's 76.0) and 94.7 on AIME 2026 reasoning.
CEO Mark Zuckerberg positioned the release as a challenge to Chinese open-source competitors (DeepSeek, Alibaba) and called for U.S. policy changes around data use and distillation. For architects building agent infrastructure, Glimmer's on-device capability matters: it enables private, low-latency agentic workflows without sending data to the cloud, and the open-weight release lowers the bar for fine-tuning and deploying custom agents. The distillation pipeline (Spark → Glimmer) is now a reference pattern for turning frontier models into deployable agents.