Meta put Muse Spark Open on Hugging Face and OpenRouter on Thursday, October 15: 816 billion parameters, 42 billion active, a context window of roughly a million tokens, listed at $1.25 per million tokens in and $4.25 out. By the small hours of Friday morning, Artificial Analysis had scored it 39 on its Intelligence Index and conferred the title that every comparison guide and newsletter then repeated verbatim — the No. 1 open-weights model in the United States.

Read the title again. Not the best open-weights model. The best American one.

‘Open’ Is Now a Distribution Channel

The word “open” has been doing a lot of unpaid labor in this industry for three years, and Muse Spark Open is where the bill comes due. The weights are downloadable. The license is not Apache, not MIT, not anything the Open Source Initiative would recognize — it’s the Muse Model License, a bespoke document written by the party that also sells the hosted endpoint. And at 816 billion parameters with 42 billion active, this is not a model anyone is running on a workstation. It’s a mixture-of-experts built for a datacenter, which means the route most teams will actually take to it is OpenRouter, at $1.25 a million tokens in.

So the artifact that the entire open-versus-closed argument has been about — the weights file — is simultaneously too big to self-host economically and governed by terms the releasing company wrote for itself. The proliferation scare and the freedom rally are both pointed at an object that functions as a product SKU. If you want to know what a lab intends, don’t read the parameter count. Read the license, then read the price list.

The Price Tag Is the Announcement

Meta’s own numbers make the competitive picture plain. Its closed flagship, Muse Spark 1.2, sits at 40 on the same index. It opened the weights to a model one point behind it. That is not a conversion experience; that is a flanking maneuver — give away last quarter’s frontier, keep this quarter’s, and use the giveaway to drag down the price of everyone serving beneath you.

The obvious target is Z.ai, whose GLM-5.3 shipped August 14 at roughly 753 billion parameters and lists at $1.40 in and $4.40 out. Meta now undercuts it by fifteen cents on input and fifteen on output, with a comparable context window and native multimodality that the Chinese model doesn’t match. That is the actual news from Thursday: not a research milestone, a price cut with a Hugging Face link attached.

Americans should be pleased about that. It’s a real competitive move by a real American lab against a real Chinese rival. It just isn’t what the headline said.

A 24 Percent Error Rate Counts as Leadership

Then there’s the number buried under the 39. Artificial Analysis noted that Muse Spark Open led open-weights models on its AA-Omniscience benchmark, “driven by the lowest hallucination rate of any open-weights model scoring 30 or more (24%).” Twenty-four percent. The best open model in the country gets roughly one in four answers wrong on the hardest knowledge test we have, and we have arranged the marketing so that this is the good news.

That’s the figure a chief information officer should be staring at, not the index score. The 39 is a composite; the 24 is what happens in production. And the open-weights framing conveniently distributes the liability — Meta didn’t ship a product with a warranty, it shipped a file.

A platform engineer at a mid-sized claims-processing software vendor, writing in her company’s internal Slack channel, put the real-world use of Thursday’s release better than any analyst has: “We’re not deploying it. We’re pricing against it. Our CFO now has a number to hold our incumbent vendor to, and that number is worth more to us than the model is.”

Which is the honest function of an open-weights release in 2026: procurement leverage for buyers, a commodity hedge for the seller. Not a public good, not an existential hazard.

The Lever Nobody Wants to Pull

If that’s right, then both sides of the open-weights debate are pulling the wrong lever, and the people drafting AI policy memos in Washington are pulling the wrongest one of all. The freedom caucus wants the weights to be free, as if the weights were the thing being governed. The safety caucus wants the weights locked down, as if an 816-billion-parameter mixture-of-experts were going to escape into a dorm room. Neither is where the constraint actually lives. It lives in the Muse Model License, in the serving layer where OpenRouter does the metering, and in whether anyone discloses the conditions under which the benchmark numbers were produced.

Meanwhile the “American open-weights leadership” crowd should be honest about what it’s celebrating. A title with a nationality baked into it is what you invent once you’ve stopped winning the unqualified one. Meta’s move on Thursday was aggressive, well-priced, and good for American buyers.

The trophy says United States. The leaderboard doesn’t have that column.

Sources