Meta Muse Glimmer landed today as a free, open weight AI model you can download and run on your own computer instead of Meta’s servers. Mark Zuckerberg announced the release Monday, alongside a promise that an open version of the more powerful Muse Spark 1.2 is coming soon. The model runs under an Apache 2.0 license, which means it is free to use, modify, and even build products on top of, but running it locally does not mean it runs on anything with a power button.

Muse Glimmer is a 30 billion parameter model from Meta Superintelligence Labs, the research group Alexandr Wang now leads inside Meta. It is built for what Meta calls always on agent workflows, tasks like managing a schedule, calling tools, checking its own results, and recovering when something goes wrong mid task. That is a different design goal from a chatbot you type questions into. Meta positions it as the engine behind an assistant that works in the background rather than one you interrupt with a prompt.

What Can You Actually Build With It

The most common assumption is that a new AI model release means image or video generation. Muse Glimmer does neither. It is a text and reasoning model that can read images as input, screenshots, charts, documents, but it does not create them. Meta has a separate model for that called Muse Image, and that is a different release entirely.

Where Glimmer holds up is coding and building. It scores well on benchmarks that test writing and debugging real code across multi step tasks, so it works as a local coding assistant running entirely on your own machine. It is also built for tool calling, meaning it can trigger external functions or scripts, check whether the result worked, and retry if something fails, the kind of thing that would sit inside an automation workflow rather than answer trivia. Add in multi step reasoning for research style tasks, and the general picture is a private, offline assistant for coding, automation, and background work, not a free image generator.

The headline feature is that it fits on a single consumer GPU or a Mac instead of needing a data center. At full precision, a model this size would need over 55GB of memory, more than any consumer machine offers. Meta compresses the weights down to roughly 4 bit precision, shrinking the model to under 20GB, which leaves enough room for everything else it needs inside a 24 to 32GB memory budget.

That number matters more than it sounds. A MacBook Air with an M3 chip and 16GB of unified memory, a common everyday setup, falls short of even Meta’s smallest recommended quantized version, which still needs somewhere around 18 to 20GB to run properly. It is not that the model refuses to load on smaller machines. It is that memory this tight forces constant swapping to disk, turning what should be a fast local assistant into something closer to unusable. If you are wondering whether your own laptop can handle it, 16GB is roughly the line under which you are fighting the hardware more than using the model.

On paper, Muse Glimmer performs well for its size, leading competing open models on benchmarks that test multi step task completion and tool use, though it trails on tasks that involve operating a computer screen directly. Meta also says the model does not meet its own internal threshold for a frontier AI system, and rates its cybersecurity and biological risk as moderate or lower.

None of this changes what today’s release actually is for most people. It is not a product you open an app and use. It is a set of files developers and hobbyists can download and run through tools like llama.cpp or Ollama, then build on top of. Meta is back to shipping open weight models after pausing releases while it rebuilt its AI research org, and it is doing so alongside a public commitment to release something even more capable soon.

FAQ

Is Muse Glimmer free to use?
Yes. It is released under an Apache 2.0 license, which allows personal use, modification, and commercial use without paying Meta anything.

Can Muse Glimmer generate images or videos?
No. It can read images as input for tasks like interpreting a screenshot, but image and video generation comes from a separate Meta model called Muse Image.

If you want the next open weight release explained the same way, plain language, no jargon, no hype, that is what the newsletter is for. Sign up at newsletter.augustdispatch.com and get it the day it happens, not a week later.

What would you actually build with a model like this if it ran comfortably on your own machine? Curious what people are picturing.