Muse Glimmer: 30B Open-Weight Agent Model for Local Use

A 30 billion parameter model built to run locally

Meta released Muse Glimmer on August 10, 2026, describing it as a 30 billion parameter model optimized for always-on local agent workflows. Weights are released under the Apache 2.0 license and, quantized to 4-bit precision, fit under 20GB, small enough to run on a Mac or a single consumer GPU with 24 to 32GB of VRAM. [1]

Meta positions the model for end-to-end agentic task completion: tool use, multi-step reasoning and failure recovery, with text and image input across more than 100 languages. It ships with a speculative-decoding drafter Meta calls DFlash, which it says speeds up generation by 1.5 to 3.1 times depending on hardware, and Meta reports it performing competitively against comparably sized open models such as Gemma4-31B and Qwen3.6-27B. [1]

Glimmer was trained by logit distillation from, Meta's larger closed flagship model that debuted in April 2026. Spark's own weights are not released; only the distilled Glimmer is open. [1][2]

Where Meta draws the line on what it gives away

Atlas interpretation: TechCrunch frames Glimmer against Mark Zuckerberg's stated vision of distributing what he calls personal superintelligence broadly, agents Zuckerberg has said will work continuously on a person's behalf across relationships, health, career, finances and other parts of daily life, offered at free or affordable access. Releasing a capable but smaller distilled model as open weights while keeping the teacher model, Spark, closed is a concrete instance of that framing rather than a break from it: Meta gets to point to an open release supporting the broad-access story while still controlling and monetizing its most capable model. [2]

This split follows Meta's own trajectory during 2026: Muse Spark itself launched as a closed model in April, and by July Meta had put a metered Model API in front of Spark 1.1 rather than releasing its weights. Glimmer is Meta's first open-weight release since that shift, but it is a distillation of the closed line rather than a reversal of it. [2]