AI SYSTEMS

The licence that finally says yes.

Meta published a 30 billion parameter multimodal model today, small enough for one consumer card. The size will get the headlines. The three words that matter for paid work are Apache two point zero.

THE SHORT VERSION

A capable model you can run yourself is interesting. A capable model you can run yourself and bill for is a different category.

•  Meta published Muse Glimmer today, a 30 billion parameter multimodal model distilled from Muse, with weights on Hugging Face.

•  It is released under Apache 2.0, which permits commercial use without a separate agreement. That is the sentence worth reading twice.

•  It is built to run locally, described as fitting a single consumer card, with day-zero support in transformers, llama.cpp and vLLM.

•  Honest scope: this is a reasoning and agent model, aimed at coding, document work and assistants. It is not a picture generator, and nothing here replaces the model that renders your images.

Twice in the last fortnight this newsletter has told you to go and read a model card before letting anything near paid work, and both times the reason was the same: the licence was either unstated or quietly restrictive. It is worth marking the day one arrives where the answer is simply printed on the tin.

The parameter count is the part that will be quoted everywhere today. The licence is the part that changes what you are allowed to invoice for, and almost nobody leads with it.

The announcement is a company post on Hugging Face, carried the same day by several outlets, so this is a vendor release with independent coverage rather than a claim standing on its own. Muse Glimmer is described as 30 billion parameters, distilled from the larger Muse, multimodal, and intended for local agentic work. Weights and developer documentation are downloadable now.

The named uses are privacy-sensitive ones: coding, document analysis, personal assistants, and evaluation work where a model judges output. The hardware claim is a single consumer card, and support in transformers, llama.cpp and vLLM landed on day zero, which means the tooling most people already have will load it.

What is not claimed matters too. Nothing in the release says this renders images or video, and a multimodal model that reads pictures is a different animal from one that makes them. Treat it as something that can look at your work, not something that produces it.

The phrase open weights has been doing a lot of quiet work lately. It tells you the file is downloadable. It tells you nothing about whether you may use the result on a job that pays.

There are broadly three positions a model can sit in, and only the first is simple. A permissive licence such as Apache 2.0 lets you download, run, modify and use it on work that pays, with attribution and no separate agreement. A restricted open-weight licence lets you download and experiment but bars paid work, or gates it behind a paid tier, which is where several popular image models sit. And a hosted model gives you no weights at all, only terms of service that can change between the quote and the delivery.

The middle position is the one that catches people, because the download works, the output looks fine, and nothing warns you until a client asks who owns what.

This takes about ten minutes per tool and you only do it once, then write the answer down somewhere you will find it again.

First, name the licence, not the vibe. Find the actual file or terms page and read the word. Second, ask whether commercial use is permitted outright or only on a paid tier, because those are sold with the same enthusiasm. Third, ask what attribution is required and where it has to appear, since a requirement you cannot satisfy inside a client deliverable is a no dressed as a yes. Fourth, ask what happens to the work you already delivered if the terms change next quarter.

A permissive licence answers all four at once, which is the whole reason today is worth noting. Most releases do not, and the ones that stay quiet about it are quiet on purpose.

Paste this, then paste the licence or terms text underneath it. It is written to refuse to guess, which is the only useful behaviour here.

Below is the licence or terms of service for a tool I am considering using on paid client work. Answer only these four questions, in plain words, quoting the exact line you relied on for each. One, am I permitted to use the output commercially, yes or no or unclear. Two, is that permission unconditional or does it depend on a paid tier, a revenue limit, or company size. Three, what attribution is required and where must it appear. Four, does the document reserve the right to change these terms, and does any change apply to work already delivered. If the text does not answer a question, write not stated and stop. Do not infer from what is customary, do not tell me what similar licences usually say, and do not give legal advice.

The instruction to quote the line is what makes this safe to rely on. Without it you get a confident summary of what that kind of licence normally says, which is exactly the failure you were trying to avoid.

THE NEXT WORKING NOTE

Still building the licence sheet: every model and tool I actually use, with its licence named, whether paid work is allowed, and the attribution line copied out ready to paste into a deliverable.

Want it when it ships? Reply with send me the licence sheet and I will get it to you.

A QUESTION FOR YOU

Has a client ever asked you what licence your tools carry?

Reply with what happened. I suspect the answer is mostly no and that it is starting to change, and if it is changing for you I would rather write about that than guess at it.

If this was useful, forward it to whoever signs off your contracts.