Zuck Buys a Round

Mark Zuckerberg cannot drink a Guinness. I know because, like a great many other people this week, I watched him try.

This afternoon I sat in a cinema with my daughter watching The Social Reckoning, Aaron Sorkin’s second attempt at the man, and came out disappointed. The Social Network is one of the twenty best films I have seen. This one pulls its punches. It reaches for the storming of the Capitol as the great harm Facebook did to the world, and that is not the harm. The harm is the children who stopped liking themselves, the ones who hurt themselves, the ones who are no longer here, and the years in which the company knew and chose growth. That is harder to put on a screen. It is also the true story, and in August Meta agreed to pay up to $17.1 billion to fifty-one attorneys general rather than have it told in court.

On the way home, there he was on my phone, pint in hand, his face running a search for the correct response to a liquid and returning no results.

And last week the same man gave the world a present: Muse Spark Open, 816 billion parameters of artificial intelligence, open-weight, free to download and free to change.

Three Zucks in four days, and together they pose one question:

A company that chose growth over children, run by a man who cannot process a beer, owns a measurable share of the planet’s attention and trades on it every day. When that company gives something away, why?

The only sentence worth reading is the one with the asterisk. So I went looking for it, and the first thing I found was that I didn’t understand the word on the label.

I have done this once before. Last year I discovered I had spent two years telling people what GPT stood for, and had been wrong, so I looked it up and wrote down what I found. Same method this time. Here is what the label means, stripped of the marketing.

An AI model is, underneath everything, a vast number of dials. Eight hundred and sixteen billion of them, in this case. Training is the process of setting every dial to the position that makes the model’s answers least wrong, and it is done once, at enormous cost, before you ever meet it. Those dial positions are the weights.

Open weight means you are given the dial positions. You can copy them onto your own machine, run them without asking anyone’s permission, and, if you have the hardware and the appetite, nudge them.

It does not mean you are given the recipe. The data the model was trained on, the code that trained it, the method, the choices: none of that is published. Meta’s own documentation states no training-token count, no compute figure and no knowledge cutoff. You get the cake, boxed, in a van. You do not get the kitchen. The cake is yours to ice however you like.

That is the difference between open weight and open source. Under the definition the Open Source Initiative published for AI, a model released this way doesn’t qualify: open source would mean you could rebuild it. Open weight means you can taste it.

And “changing the weights” is less dramatic than it sounds. Nobody edits eight hundred billion dials by hand. You either train the model a little further on your own material, or bolt a thin layer on top and leave the original untouched, which is what almost everyone does.

For two hundred years the British brewing trade ran on an arrangement called the tied house. The brewery found a landlord, gave him the pub on an easy lease, lent him the money to fit it out, and asked for one thing in return: he bought his beer from them, at their price, for as long as he held the keys. The pub was cheap. The beer never was. The trade had a name for the difference, and it was “wet rent.”

The licence attached to Muse Spark Open, which Meta calls the Muse Model License, is a tied house, clause for clause.

The pub, on easy terms. Anyone may download the weights, run them, modify them, build on them and redistribute them, for nothing. The beer is everything you pour through it afterwards: the sign you carry, the terms Meta can change, and the architecture your people learn to work in. That is the wet rent, and it is paid for as long as you hold the keys.

Our sign over the door. The licence requires that every product built with the model “prominently display ‘Built with Muse Spark Open’ in the user interface.” Intel Inside, for AI, written by a company that understands what a sign is worth.

No rival brewery. If you provide “access to AI models as a service” and your group’s gross revenue “exceeds twenty (20) million US dollars” in any twelve months, you must notify Meta and sign a separate agreement before you serve its beer. That clause is not aimed at you. It is aimed at the clouds and the hosts who would otherwise make a business of selling Meta’s model back to the world. Llama, this model’s predecessor, carried a gate of its own, pitched as I recall at companies with more than 700 million users; this one is measured in revenue instead, and the bar is far lower.

And the cellar stays locked. The model Meta gave away is not the best model Meta has. Three days earlier it launched Muse Wonder, which it says is at the frontier, and kept it closed. Alexandr Wang, Meta’s Chief AI Officer: “Because of both safety and capabilities, we will generally not open-source our most powerful models.” Look at the whole family and you can see the ladder. Muse Glimmer, the small one, released in August under a licence as close to free as licences get. Spark Open, the big one, under the tied-house terms. Wonder, the best one, behind the bar. The smaller the model, the freer the licence.

The remaining asterisks, in the order a buyer meets them.

Before you may download the free thing, you fill in a form: legal name, date of birth, organisation. A company that has just paid seventeen billion dollars over what it did with young people’s data and attention would like to know who you are before it gives you anything.

The licence says the model “is not intended for individuals under the age of 18.” A company whose products were built to hold a thirteen-year-old’s attention has written into its licence that this one isn’t for children. Read that twice.

The sign is theirs. Whatever you build, their name sits on it.

And the ingredients cannot be inspected. If you care what trained the thing that is about to write your copy, draft your research or talk to your customers, this is the asterisk that matters, and no amount of “open” in the name removes it.

One honesty note. “The most capable US open-weight model to date” is Meta’s phrase, and the small asterisk is on “US.” Meta’s own comparison table shows three Chinese models, GLM 5.3, Kimi K3 and Qwen3.8, ahead of it on several of the coding and agent tasks it was built for.

So why does a company that has just paid seventeen billion dollars to keep its own conduct out of a courtroom give a thing like this away? Four reasons, and generosity isn’t one of them.

One. Meta is not in the business of selling AI. It is in the business of attention, and of the advertising and commerce that attention is converted into. Google did not give Android away because it wanted to sell phones. It gave Android away so that nobody else could own the layer between Google and its users. Cheap, everywhere AI costs Meta nothing it was ever going to charge for, and denies the companies that do sell AI the chance to own the interface through which Meta’s users will reach it.

Two. The estate. This is the wet rent being collected. Within a day of release the model was running in every major inference engine, on Apple’s own chips, with a published fine-tuning recipe from Hugging Face. Every developer who builds on it is working, unpaid, on Meta’s architecture. The reading I’d add, as mine rather than Meta’s, is that work done on the open model flows back into the closed ones and into the enormous compute Meta is building for itself; weigh it as inference.

Three. The licence is the moat around the gift. The twenty-million-dollar clause means the only companies that could turn Meta’s model into a rival business must first sit down with Meta.

Four. The flag. Wang’s post announcing the model closed, in lower case, with “america!” Meta’s own title for the release was “the most capable US open-weight model to date.” A free model is also a lobbying position. Last month I wrote about the frontier labs asking, in chorus, to be regulated. Meta has chosen the other pew: not asking to be regulated, asking to be the standard.

I keep coming back to the face. A man so far outside ordinary experience that a pint defeats him, who has spent twenty years deciding what three billion people see, has just bought the whole house a round. The round is real. The pub is yours for nothing. And above the door, where you might not have thought to look, hangs a sign with his name on it, and every pint you pour from here on is bought on his terms.

The pub is free. The beer never is.

So here is my question, and it’s a practical one. What’s the “free” thing in your stack right now and whose sign is over the door?

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry’s table: senior client, global agency leader, technology founder… three different bars at which to be handed the tab. Free, every Monday and Thursday, with no sign over your door: brandflow-marketing.com.