The best open weight AI model now runs on a desktop.
Yes, you can run the best open AI models at home. I timed it. Five words a minute.
Two weeks ago I wrote that AI forces a choice: open, capable or sovereign, pick two. The near frontier open weight models are too big for almost anyone to operate.
The pushback was consistent. You can run these models yourself.
Running a model and running your business on it are two different things. So let me put numbers on that.
One project doing exactly this is Colibrì, fully open source under the Apache 2.0 license. I came across it when it launched at the start of July.
The day after I published that post, it shipped support for Kimi K3, the Chinese model that beats the American leaders at writing code. The release note was blunt: "nobody has run it end to end yet." So I went back and looked properly.
LLMs are enormous, but use only a small fraction of themselves for any single word. So Colibrì keeps that part in working memory and leaves the rest on an SSD, pulling pieces off the drive as the model calls for them.
Which means the hard drive becomes the engine. And the speed of that drive becomes the speed of your AI.
It works. Kimi K3 runs on a desktop, and here is the cost.
The model takes 1.5 terabytes of storage. The project's own measurement, after heavy tuning of how it reads that drive, is 9.4 seconds per token, the small chunks of text these models produce. A one-page answer takes over an hour.
I tested the mid-sized model, GLM-5.2, on my own machine. Startup was brutal. Once it settled it ran at roughly 1.8 tokens a second, slower than someone reading aloud.
And it gets worse. That is one person running one job. Not a dozen colleagues with sessions open at once.
If you work alone, that might be enough. For an organization it is not a starting point.
The developer who built Colibrì runs it on six RTX 5090 graphics cards, far past any sane budget, and gets 5.8 to 9.2 tokens a second. A slow commercial provider running the same model is five times faster, and serves thousands of customers at once.
For now, what you own is a fire escape. Nobody judges a fire escape by how fast it gets you to the lobby.
Which leaves anyone serious about this with two doors. Build inference capacity on your own premises, a capital investment and the people to run it. Or rent again, from a provider who then holds your data and needs the same hard questions about whose law reaches them.
There is no third state. A dependency is either working or it is gone. The escape hatch people keep pointing to does not survive a working week.
None of which takes anything away from the work itself. Streaming a model off a drive was not on anyone's roadmap a year ago. It is not production ready, and the numbers above say so plainly. But the direction is right, and it is being set in the open.
If you have built anything for sovereign AI, what does it actually look like, and what did it cost you to find out?
