aiexpert
Home / News / Brief
Breaking · Aug 11, 2026, 12:02 AM · 4 sources

Meta releases Muse Glimmer: 30B open-weight model for local agents, distilled from Muse Spark, runs on laptops

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.

Sources

Everything this brief rests on
  1. 01 Primary source thenewstack.io
  2. 02 cnbc.com cnbc.com “Zuckerberg said that the U.S. will need to 'rethink' its policies in several areas, including distillation and data use in training, if it is to lead in open source AI”
  3. 03 39.quartz.com 39.quartz.com “The model is a distilled version of Muse Spark 1.2... Meta trained Muse Glimmer by transferring reasoning capabilities from that larger model, and added a lightweight companion model to speed up text generation”
  4. 04 research.meta.ai research.meta.ai “Muse Glimmer is a 30-billion-parameter open agentic model from Meta Superintelligence Labs, optimized for always-on local workflows on consumer hardware”