Open for Whom?

Meta released Muse Spark Open on October 15 — 816 billion parameters, 42 billion active at a time, a million tokens of context — and most of the coverage I saw led with the same detail: it runs on two Mac Studios.

I wanted that to be true in a useful way. A frontier-class model a normal team could run under a desk would change real decisions for people like me — no API bill, no data leaving the building, no vendor in the middle. So I looked at what “two Mac Studios” actually means, and that’s where the story started to come apart.

The machines in the demo are M5 Ultra Mac Studios with 512GB of memory each. That configuration wasn’t shipping when Meta made the announcement — Apple has it coming later this month, price unannounced. The version you could actually order was the 256GB machine, at about $10,799 apiece, and since 256GB isn’t enough on its own, the working setup takes four of them linked together. Call it $38,000 to $43,000 worth of computers before you’ve served a single request.

What you get for that is a demo, not a deployment. Apple’s own numbers put it at roughly 25 tokens a second on the two-machine setup, 37 on four machines — readable output, not a service you’d put a team on. The weights explain why: 1.65 terabytes at full precision, around 537GB even at 4-bit, and the whole model has to live in memory whether or not a given part of it is firing. The compute is cheap. The memory isn’t.

None of this makes Spark Open a bad model. It makes the headline misleading. In practice, almost everyone will use it through somebody else’s API — OpenRouter had it at about $1.25 per million input tokens and $4.25 per million output at launch, ordinary cloud pricing. Fine. But then it’s a hosted model with downloadable weights, a much less interesting announcement than the one that got made.

The license bothered me more, once I read it. Spark Open doesn’t ship under an open source license. It’s Meta’s own Muse Model License. Products built on it have to say “Built with Muse Spark Open,” and if you’re offering the model itself as an API service and your company does more than $20 million in revenue, you have to notify Meta and sign a separate agreement first.

I should be fair about that clause, because most of the commentary I read wasn’t. It exempts internal use, research, and using the model as a feature inside your own product. If you’re doing what my teams would do — running agents for your own engineers — it almost certainly never touches you. It’s aimed at people reselling access. But it’s still Meta’s license, and terms like that can be changed and enforced by the company that wrote them, and you don’t get a vote. I’ve sat through enough vendor negotiations to know the difference between building on Apache 2.0 and building on a license named after the vendor. One is a settled fact. The other is a relationship, and relationships with platform companies get renegotiated on their schedule, not yours.

The contrast showed up the next day. On October 16, Reflection released Beam — 501 billion parameters, similar context window, weights under plain Apache 2.0. No attribution requirement, no revenue threshold, nothing to sign. They claim 81.6% on SWE-bench Verified — their number, not an independent one, and I wouldn’t plan around it until somebody else reproduces it.

Beam is what an open license is supposed to look like. What it isn’t is cheap to run: even at 4-bit it’s roughly 250 to 280GB of weights — a rack of server GPUs, not a desktop, and not a line item most small or mid-sized companies will approve so they can say they self-host. So in two days we got a restrictive license almost nobody can run locally, and a clean license almost nobody can run locally either. Different problems, same result for most teams.

The irony is that Meta already ships the model that fits the original headline: Muse Glimmer, 30 billion parameters, Apache 2.0, released in August, built to run on a single consumer GPU. A small team could have it running this week, on a machine they already own, under a license they don’t need a lawyer to parse. It got almost no attention, presumably because “mid-sized model runs on hardware you have” is not much of a keynote.

I don’t think open weights are a scam — releasing them at this scale is real work, and both models will get used, mostly through APIs, sometimes valuably. The lesson is dumber and more practical: the adjective in the announcement is doing a job, and the job is not informing you. So when the next one lands, I ask the questions nobody put in the press release. What does it cost to run, on machines I can buy this quarter, at a speed my team would tolerate? And what does the license actually say — the whole thing, including the parts that only apply once we’re successful enough to be worth noticing? If I can’t answer both from primary sources, I don’t have an open model. I have a marketing claim with a download link.

“Open” is a good word, and it used to mean something specific. I’d like it to still mean the people it’s supposedly for can actually use the thing — a team like mine, on a budget like mine, without a hardware lab or a legal department clearing the way first. Spark Open isn’t open for us. Beam is closer, and it’s still not close.

Figures as reported Oct 15–19, 2026. I’ve read the license summaries in the press coverage, not had counsel review the license itself, and Beam’s benchmark is vendor-reported and not yet independently replicated.