The Word Open Was Doing Two Jobs

2026-07-29


๐Ÿ”“ The Word Open Was Doing Two Jobs July 28, 2026 ยท https://tavi-blog.github.io/the-word-open-was-doing-two-jobs/

I checked, out of habit more than hope, whether the model everyone spent this week calling the most open release in the field would run on the same rig I use to run the ones that actually do. It wouldn't. Not close. The weights alone come in at over half a terabyte in their compressed form, and the number Moonshot itself publishes for a workable setup starts at eight top-tier accelerators wired together, the kind of cluster a research lab rents by the hour and a person does not own. I closed the tab and went back to the small model already running locally on this laptop, the one I picked specifically because I could.

The release itself is a real thing and deserves to be taken at face value before I get to the part that bothers me. A lab put out a 2.8-trillion-parameter model under a fully permissive license, no restriction on commercial use, no fine-tuning clause to negotiate around, full weights on a public repository instead of a metered API. Coverage this week has been calling it a shot across the bow of the closed labs, and there's something to that framing. A commercial team building a product no longer has to accept whatever pricing and rate limits a closed provider sets, or wonder whether a model they depend on gets deprecated out from under them on someone else's schedule. That is a genuine transfer of leverage away from three or four companies and toward anyone with the infrastructure budget to use it, and I don't want to undersell how real that shift is for the businesses it actually reaches.

But that phrase, anyone with the infrastructure budget, is doing almost all the work in that sentence, and it's the part that gets quietly dropped when this gets described as an open release in the same breath as the small self-hostable models I actually run. I got into local models for a specific reason: I wanted to point something at my own files without first deciding whether I trusted a company with them, on hardware I already owned, no subscription, no API key, no data leaving the machine. I built an eGPU rig around a modest consumer card partly so a single laptop could do that and still play a game at night, one machine instead of two, because I don't want a second device to charge and update just to have a model that answers to me instead of a vendor. That's the actual lived version of "open source as a value," and it has a floor. It has always assumed a person could plausibly clear that floor with a weekend and a decent GPU.

A model whose minimum viable footprint is a rack of enterprise accelerators clears a completely different bar. It's open in the sense that matters to a company: no license fee, no lock-in, full audit rights over the weights. It is not open in the sense that matters to me, because I was never going to be the one running it regardless of what the license says. The two things used to travel together, more or less, back when an "open" model meant something you could realistically pull down and run yourself if you cared enough to bother. Somewhere in the last year the frontier of open-weight releases quietly outgrew that assumption, and the word kept getting used as if nothing had changed, because from a boardroom the distinction between "anyone can run this" and "any company big enough can run this" doesn't look like much of a distinction at all.

I don't think the people releasing these weights are being dishonest about what they built. Getting a model this capable out from behind a paywall, even one only a well-funded team can operate, is still a different world than the one where three companies hold all the good weights and rent access by the token. I'd rather that leverage sit with more labs than fewer. I just don't think it's the same win I was cheering for when I first started choosing tools I could run over tools I had to trust, and I'd rather notice the difference now than keep filing both under the same word until I've forgotten there were ever two.

The small model still answers fewer questions than I'd like and gets more of them wrong than the one I'll never touch. Whatever the license on that other one says about who's allowed to call it open, the only model on this laptop that answers to me instead of a data center is the one I already had.


Don't miss what's next. Subscribe to tavi-blog: