Muse Glimmer is the open-weight entry in Meta’s developer line-up, and that single word changes the relationship. Everything else on this site is an app that needs an account, a network and your permission to look at things. A model whose weights you have downloaded needs none of them and hands you a set of obligations the app was quietly carrying on your behalf.
A model’s weights are the numbers that came out of training billions of parameters, and nothing else. They are enough to run the model and to adapt it. They are not enough to understand how it came to be, and they are not a licence.
The analogy that fits is a baked cake rather than a recipe. You can eat it, slice it, decorate it and work out a fair amount about it by tasting. You cannot reproduce it, you cannot see which ingredients went in, and you cannot check whether anything in it disagrees with you.
That gap is why “open-weight” caught on as a separate term. It was coined precisely because calling these releases open source was inaccurate, and the distinction has teeth.
It matters practically, not just pedantically. Without the training data you cannot audit what the model learned; without the training code you cannot reproduce or verify the result; and without reading the licence you do not know what you are allowed to do with any of it.
None of which makes open weights a lesser thing. Being able to download a capable model and run it on hardware you own is a genuine shift in who holds the power in the relationship. It is simply a narrower claim than the one people hear.
Open source is a claim about rights to study, modify and redistribute. Open weights describes what was shipped. A release can be open-weight and carry a licence that forbids most of what people mean by open.
No download fee is not a grant of permission. Open-weight licences across the industry range from genuinely permissive to tightly conditional, including limits on commercial use, on scale and on what you may build.
Weights are billions of numbers. They tell you nothing readable about what the model absorbed, what it was tuned to refuse, or whose material went into it. Having them is not the same as being able to inspect them.
Running on your own machine protects your data, not your output. Whatever guardrails a hosted service applies around a model are part of the service, and a raw model on your desk arrives without them.
This is the part that gets skipped, and the only part that can put you in a bad position. A file transferring successfully tells you nothing about what you are entitled to do with it, and “it was a public download” has never been a defence.
Meta’s licence for Muse Glimmer is not reproduced or summarised here, deliberately. A paraphrase of a licence is worth nothing, and a wrong paraphrase is worse than nothing read the actual document, and for anything commercial have somebody whose job it is read it too.
The same caution applies to where the weights come from. An installer can be checked against a publisher’s signature; a set of parameters cannot be read, so weights from an unofficial mirror are trusting a stranger in a way an app download is not.
Every other page on this site describes an assistant that lives on Meta’s servers and reaches into your machine. Weights you hold invert that arrangement completely, and the consequences run further than privacy.
| The Meta AI Mac app | Weights on your machine | |
|---|---|---|
| Where it runs | Meta’s servers; the app is the window. | Your Mac, using its own memory and chip. |
| What leaves the machine | Your prompts, and any window you attach. | Nothing, once the weights are downloaded. |
| What it needs from you | A Meta account and system permissions. | Disk space, memory and patience. |
| Network | Required. No connection, no assistant. | Optional. It runs on a plane. |
| Cost shape | Free with limits on heavy use; a subscription raises them. | No per-request cost. Hardware and electricity instead. |
| When Meta changes it | It changes for you, whether or not you wanted it to. | It does not. Your copy is your copy. |
Privacy gets the attention, but permanence is what people notice months later. A hosted model can be updated, restricted, repriced or retired, and the change arrives whether you asked for it or not. A downloaded model behaves the same way next year as it does today, which is worth a great deal if you have built something on top of its exact behaviour.
Whether any particular Mac can run any particular model depends on the model’s size against your machine’s memory. Apple silicon is unusually well suited to this because the chip and memory are shared rather than split across a separate graphics card but the size figure has to come from Meta, not from here.
A hosted app is a managed service, and most of what it manages is invisible until you stop paying somebody to do it. These four transfer to you the moment you run weights yourself.
Memory is the binding constraint, and a model either fits or it does not. Where it fits but only just, everything runs slowly enough to change how you use it.
Previously handled by: whoever provisioned Meta’s servers. The free tier’s usage limits were the visible edge of a cost somebody else was absorbing.
Something has to load the weights, serve requests and keep running. On one machine for one person that is mild; for anything shared it is a real service with real uptime expectations.
Previously handled by: the app, which reconnected and retried without telling you it had.
Hosted assistants apply filtering and policy around the model. A raw model has whatever behaviour training gave it and nothing sitting in front of it, which matters most if you are putting its output in front of other people.
Previously handled by: Meta, as part of the service and it is the transfer people are least likely to have planned for.
Nobody will tell you the model got worse at your task, because nobody is watching your task. If you swap models or change how you run one, measuring the difference is on you.
Previously handled by: nobody, in fairness but a hosted service at least published figures you could argue with.
The reasoning behind that split, and why this site carries no model identifiers or quotas, is set out on the Meta Model API page.
You can download the trained parameters and run the model yourself. You do not necessarily get the training data, the training code or the recipe that produced it, which is why open-weight and open-source are different claims. The distinction in full.
No. Open source is a statement about licensing and about being able to study, modify and redistribute the thing. Open weights describes what was released the parameters and says nothing on its own about what the licence permits.
That is a question only the licence answers, and it is a separate document from the download. Open-weight releases across the industry vary from permissive to heavily restricted, so never infer permission from the fact that a file downloaded successfully. Why this matters.
Whether a given Mac can run a given model depends on the model’s size against your machine’s memory. Apple silicon’s unified memory makes Macs unusually capable at this for their price, but the size figure has to come from Meta’s own documentation rather than from a page like this one.
Nothing leaves your machine, there is no account, no per-request cost and no network dependency and nobody can change or withdraw the model under you. A downloaded model is a copy you keep, which is a different relationship from a service you use. Six rows of difference.
You become the infrastructure. Hardware, serving, updates, evaluation and safety behaviour all become yours to handle, and none of them are handled for you the way they are in a hosted app. The four jobs.
Not automatically. There may be no fee to download, but the licence can still restrict use, and running a model has real costs in hardware and electricity that a hosted service quietly absorbs on your behalf.
Muse Spark is the general-purpose model behind Meta AI, and Muse Glimmer is listed separately among the developer tools. Their relationship, and any difference in capability, is not something this page can state from published material.
Having the weights makes fine-tuning technically possible, but whether you are permitted to, and what you may do with the result, are licence questions. Some releases also place conditions on models derived from them.
From Meta’s own developer documentation and whatever distribution channel it names. Weights obtained from an unofficial mirror are untrustworthy in a way an app download is not, because you cannot inspect what a set of parameters will do.
It is still a real one. You get a capable model you can run on your own hardware, keep indefinitely and adapt within whatever the licence actually says.
Independent guide not affiliated with Meta and not official documentation. Model sizes, licence terms and distribution come from Meta’s developer documentation. What is actually known today.