Meta Releases Muse Glimmer, A 30B Model That Runs On A Single Consumer GPU

Meta released Muse Glimmer on Monday, a 30-billion-parameter model built for agent work that fits on a laptop. The company published the model’s weights – the trained numbers that make up the model itself, meaning anyone can download it and run it on their own machine – on the model repository Hugging Face, under an Apache 2.0 license that is genuinely permissive, without the usage restrictions Meta attached to its Llama releases.

The model is aimed at a specific and increasingly crowded target: AI that runs on your own hardware instead of somebody else’s cloud. Google has Gemma, Alibaba has Qwen, Mistral and DeepSeek both ship small open models. Meta is arriving late to a category it arguably created and then abandoned.

What It Is

Glimmer was built from Muse Spark, Meta’s frontier model, using a technique called distillation: you train the small model on the big model’s output until it learns to imitate what the larger system already knows. 

Then there’s the problem of making it fit. A 30-billion-parameter model at full precision needs more than 55GB of memory, which no consumer graphics card has. Meta compressed the numbers that make up the model down to roughly a quarter of their usual precision, shrinking it to under 20GB – small enough to leave room for everything else the model needs running alongside it inside a 24GB or 32GB card. The company says the compression costs little or nothing on the tasks that matter.

Speed comes from a second trick, called speculative decoding. Models normally write one word at a time, each one waiting on the last, which is why long answers feel slow. Meta pairs Glimmer with a small, fast companion model that predicts whole chunks of text, then has the real model check the guesses all at once and keep whatever it got right. It works because checking an answer is much faster than producing one. Meta reports the result is 3.1x faster on an RTX 5090, 1.8x on an M5 Max, and 1.5x on an M4 Max.

Meta has positioned Glimmer against Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B, and claims strong results on tests that measure whether a model can complete a multi-step job start to finish – fixing real bugs in real codebases, calling outside tools, recovering when something fails. Those are the company’s own numbers from the company’s own testing, which is worth remembering until outsiders get their hands on it. As of Monday, they can.

Meta CEO Mark Zuckerberg says a version of Spark itself will follow in the coming weeks, with larger models after. 

Models you can download are cheaper to run and easier to customize than models you rent, and the strongest downloadable ones increasingly come out of China – DeepSeek, Alibaba, Moonshot. Zuckerberg’s argument is that American labs are hobbled by training-data restrictions their foreign rivals don’t face, and that blocking foreign models is the wrong answer to that.

“US policy must reduce this additional friction if we want American open source models to lead over time,” he wrote.

He also wants distillation protected as a matter of policy – “you can learn from anything you can observe.” Glimmer is a distilled model, released the same morning, so the principle has a beneficiary.

Meanwhile

The model came wrapped in a 6,500-word essay titled “The Future is for Everyone,” arguing that advanced AI should be handed to individuals rather than concentrated in a few institutions, and that “the notion AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic.”

Meta also announced a $1 billion fund for communities hosting its data centers, a response to the local opposition that has become one of the larger obstacles to building AI infrastructure. The essay cites Richland Parish, Louisiana, where teachers received a $50,000 bonus out of the tax revenue Meta’s construction generated.