:quality(80))
A new item has started appearing on small-business AI checklists: "evaluate self-hosting an open model." We'd like to talk you out of it, and then explain why this week's open-weight news is still the best thing to happen to your software costs this year.
The download that isn't for you
On July 27, Moonshot published the weights for Kimi K3, the largest open-weight AI model ever released. Free, public, and enormous: 2.8 trillion parameters, roughly 1.4 terabytes on disk. Running it yourself means around 18 data-center GPUs, a license to comply with before you build anything commercial, and an ongoing and extraordinary operations burden.
To be blunt: this download isn’t for you. It’s not for us either, in the way people imagine. It’s for specialized providers who serve models the way utilities serve power: at scale, redundantly, with staff on call. Most enterprises with actual platform teams will consume K3 through an API rather than rack it themselves. A ten-person agency or a solo consultancy has no business anywhere near a weights file, and that is a division of labor, not a limitation.
The part that is for you
Open weights make serving competitive. Because anyone can host the same model, the companies that do host it compete on price and reliability, with positive market effects: hosted open endpoints run 60–90% cheaper than leading closed offerings for comparable everyday work, per BasedAI's research report The Open-Weight Inflection. When a bigger, better set of weights hits the market as it did this week, that competition starts again at a higher level of capability.
You’ve seen this before. Your phone, your router, and half the internet run on Linux, and you have never compiled a kernel. Open source won by disappearing into the background of products that just work. Open AI is following the same route, and this week was a large step along it.
The point isn't to host your own AI model. The real checklist item is making sure the AI tools you pay for are keeping up with this trend. They should do real work for you and be built on open-weight technology, so the cost savings end up in your pocket instead of padding someone else's profit margin.
Where the value accrues
While everyone was focused on the latest model announcements, small teams were just doing what they always do. Following up on outreach that's gone quiet. Drafting marketing content that's still half done. Chasing invoices that haven't been paid yet. Keeping a calendar that only works if you stay on top of it.
The data has been clear for months: small businesses bring in AI fastest where the work is easy to see. And it helps people rather than replacing them. According to Intuit's May survey of 34,000 small businesses, companies using AI are four times more likely to be hiring than laying people off.
That's where Kōhai comes in. Instead of juggling five separate tools, you get one team of AI coworkers. They live in your inbox, CRM, and calendar. They draft your marketing content, handle outreach follow-ups, keep invoice chasing and expense admin moving, and pause for your approval on anything that matters.
It's built on open-weight models that are served the right way. That's why weeks like this one make it better without making it more expensive. And it's why the pricing doesn't hit you with a surprise token bill at the end of the month.
What this means for you
Don't worry about the model weights. Don't bother reading benchmark tables. Just ask one question about every AI tool you pay for: when the open models get better, does my cost go down and my output improve? Or does the vendor just keep the difference?
The Kōhai closed beta is open at Kōhai.co, and we're opening to more teams soon. Describe the work; your first AI coworker handles the rest.
Kōhai is a product of BasedAI, the acceleration and commercialization layer for open source AI.
Sources: TECHi — Kimi K3's open weights arrive July 27 (July 2026) · BasedAI Research — The Open-Weight Inflection (July 2026) · Intuit QuickBooks AI Impact Report (May 2026)