I went to try the best open model in the world. It was sold out.

On Monday I sat down to do the thing I always do with a model everyone is shouting about: use it for a week on real work and report what breaks. Kimi K3, from the Chinese lab Moonshot, had spent the weekend being called the best open model ever built, topping a coding leaderboard within a day of launch and, according to my colleagues on the markets desk, nudging American chip stocks around on a Friday. I am in Singapore, a good deal closer to Moonshot than most of the people writing about it, and I had a real task lined up. I went to subscribe. I could not. New sign-ups were paused.

That is the whole review, in a sense, and it is a more useful review than the one I planned. The best open model in the world, this week, is a thing you cannot get into.

Moonshot said so itself, and I appreciated the candor. In a post over the weekend the company wrote that Kimi K3 had “received far more love than we expected, and our GPUs are feeling it,” and that demand over roughly forty-eight hours had pushed close to the limits of its current capacity. So it paused new subscriptions to protect the people already paying, and said fresh slots would reopen in batches rather than all at once. Existing users, unaffected. New users, me included, at the back of a line that does not yet have a length.

What I was actually trying to do

The task was ordinary, which is the point. I wanted to hand K3 a medium-sized, slightly boring job — take a working but ugly personal project, ask the model to refactor a chunk of it, then live with the result for a few days and see whether its changes survived contact with the rest of the code. This is the test I trust, because it is the test that separates a model that demos well from a model that helps on a Tuesday. Leaderboards measure the demo. The Tuesday is the part nobody screenshots.

I never got to the Tuesday. I got to a sign-up page and a polite message. And after the small indignity wore off, I realized the wall itself was telling me something more interesting than any benchmark could, because it exposed the quiet gap in a phrase everyone has started using without inspecting it. Kimi K3 is an “open” model. I could not open it.

“Open” is doing a lot of work in “open model”

Here is the part that trips people up, and it tripped me up for a second too. Moonshot has said it will release the weights for Kimi K3 on July 27. In the language of this industry, that makes K3 open: the actual model, the thing itself, will be a free download. And the instinct, if you have heard “open source” your whole computing life, is to assume that means you will be able to run it. Grab the file, point your machine at it, go.

You will not be able to run it. Not on a laptop, not on a gaming rig with a nice graphics card, not on a single rented cloud GPU. Kimi K3 is a 2.8-trillion-parameter model — reportedly the largest open-weight model anyone has released — and serving it takes something on the order of eight top-end data-center GPUs, the H100 or H200 class, wired together. That is roughly a quarter of a million dollars of hardware before you have paid the power bill, and that is to run one copy for a handful of users. The weights being free does not make the machine to run them free. It makes them free the way a container ship is free if someone hands you the blueprints.

The weights being free does not make the machine to run them free. Kimi K3 is open the way a container ship is open if someone hands you the blueprints.

So what does the July 27 release actually get a normal person? Honestly, not much directly. It gets a lot for a specific set of people: cloud providers who will host K3 and rent it to you by the token, enterprises with their own clusters who want to run a frontier-grade model inside their own walls, and researchers who want to poke at its insides. Those are real and they matter. But if you are a developer who read the leaderboard and wanted to try the thing this week, “open weights” is a promise made to someone else on your behalf. The practical path back to you runs through a hosted service — Moonshot’s, which is paused, or a third party’s, which will spin up over the coming weeks, priced above free, and drawing on the same scarce pool of GPUs that just made Moonshot cry uncle.

The wall is behind the license now

For years the argument about AI access was about licenses. Open weights versus closed. Could you see the model, modify it, self-host it, escape the vendor. That argument still matters, but this week made me think it has quietly stopped being the binding constraint. The binding constraint is the GPUs, and the GPU shortage is completely indifferent to whether a model is open or closed.

Think about what actually happened here. A well-funded lab — not a hobbyist, a company reportedly worth around thirty billion dollars and preparing to go public — launched its flagship model and ran out of capacity to serve it inside two days. If Moonshot, which builds the model and presumably has first claim on its own hardware, cannot keep up with demand, the idea that “open” meaningfully lowers the barrier for the rest of us needs an asterisk the size of a data center. The model is open. The compute is not. And the compute is the part you cannot download.

There is a genuine irony in it. The whole promise of an open frontier model is that it routes around the gatekeepers — no single company deciding who gets to use the best AI. But if using the best AI requires eight H100s per instance, then the gate did not disappear. It moved. It moved from a license agreement, which a lab controls, to a supply of accelerators, which almost nobody controls, and which is currently being bid up by every hyperscaler on earth. Openness got you through the first door and into a second room with a bigger lock.

Credit where it is due

I bury things for a living, so let me be fair before I finish, because there is a lot here that is genuinely good. Running out of GPUs is not a failure mode you hit with a bad model; you hit it with one people actually want, and the demand appears to be real. Topping a coding leaderboard on day one is a gameable achievement but not a meaningless one. And Moonshot’s specific choice — to pause new sign-ups rather than let quality degrade for the people already paying — is the honest one. I would rather a company tell me the line is full than let me in and quietly make the product worse for everyone. That is a decision I wish more Western services made under load instead of pretending nothing is wrong.

The open release matters too, just not for me and not this week. A 2.8-trillion-parameter model with downloadable weights is a real gift to the people who can host it, and over the next few months it will show up on the menus of the cloud providers and inference startups, and the price of using a frontier-class model will fall because there is now one more of them that anyone with a cluster can serve. That is how this is supposed to work. It is just slower, and more indirect, and more expensive at the point of use than the word “free” makes it sound.

The business context explains the squeeze, if you are wondering how a company this well-capitalized gets caught short. Moonshot has reportedly raised billions this year and is preparing a Hong Kong listing; a hit model days before an IPO is a wonderful problem and an expensive one. Serving K3 at the scale its launch demanded would mean buying or renting a lot more of exactly the hardware everyone is fighting over. Pausing sign-ups is cheaper than losing money on every new user, and it doubles as a story about demand. I do not think that is cynical. I think it is just what the arithmetic forces when the thing customers want costs more to deliver than they pay.

The verdict

So, the usual close: who is this for, and who should wait. If you run a cluster, or you have an enterprise budget and a provider ready to host K3 after the 27th, this is a model worth your attention — watch what the independent hosts charge and whether the quality holds up outside Moonshot’s own harness. If you are a normal developer or a curious user who saw the headlines and wanted to try the best open model in the world this week, wait, because you have no other choice, and know that the open release will not hand you the keys either. It will hand them to your cloud provider.

I keep a running graveyard of subscriptions I have signed up for and cancelled, and I cite it to keep myself honest. Kimi K3 is the first entry I have ever had to file under a new heading: a subscription I could not sign up for in the first place. The model may well be everything the leaderboards say. I would love to tell you whether it survives a Tuesday. Ask me after the 27th, and after somebody with more GPUs than me turns it on.