Meta Releases Muse Glimmer, a 30B Open Weight Model That Runs on One GPU

Meta Superintelligence Labs has released Muse Glimmer, a 30 billion parameter open weight model built specifically for agent tasks: planning, executing, recovering from failure, and completing multi-step jobs without constant supervision. It runs on a single consumer GPU with 24GB of VRAM, works with or without an internet connection, and is licensed under Apache 2.0.

That last detail matters as much as the parameter count. Apache 2.0 is a genuinely permissive licence, letting teams fine-tune, redistribute and build commercial products on top of the model without the usage restrictions attached to some other "open" releases. According to Meta's research blog, Glimmer decodes roughly three times faster than comparable models and works with vLLM from day one, which is the difference between a model that is technically open and one a team can actually deploy without weeks of infrastructure work first.

The target workloads are specific: function calling, local coding assistance, long tool-use sessions, and using a model as a judge to evaluate other models' output. All four are jobs that currently mean sending data to a cloud API, with the latency and privacy trade-offs that involves. A model that runs entirely on a single local GPU removes both, at the cost of whatever capability gap still exists between a 30B open model and the largest closed ones.

The release lands alongside a broader position from Meta: founder Mark Zuckerberg has been publicly critical of closed AI labs in recent weeks, and Glimmer reads as Meta putting real weight, both figuratively and in gigabytes, behind that argument rather than just making the case in an essay.

For studios and agencies experimenting with AI tooling, the practical appeal is control. A model you can run entirely offline, on hardware you already own, is one you can build workflows around without a subscription, a rate limit or a data policy someone else can change without warning. It will not replace the frontier closed models for the hardest reasoning tasks. For the narrower, repetitive agentic work most teams actually need done, coding assistance, tool use, evaluation, it is a serious option, and a sign that open enough to actually deploy is becoming a real second track in the AI market, not just a research curiosity.