Listen to this article
AI may have been used to generate this audio.
Press Space or K to play or pause, and J or L to go back or forward 15 seconds.Kimi K3 landed in July calling itself Open Frontier Intelligence. 2,107 points, 1,216 comments, and most of the argument was the usual one: Is it really frontier, does it beat the closed models, which benchmark are you quoting.
That argument is a distraction, and I say that as someone who wrote a whole post about not trusting a single benchmark score.
Open weights change the verb
With a hosted model you can use it. With open weights you can also inspect it, fine-tune it, run it where the data is, keep a version frozen for three years, and audit what changed between releases.
Those are different verbs, and only one of them lets you build something you fully own.
For anyone working outside English, or outside the domains the big labs care about, that difference is practical. If your material is unusual (a dialect, a regulated internal corpus, a domain vocabulary nobody has tokenised well), fine-tuning is the only path to a model that is good at your thing. Hosted APIs do not offer that, and they will not, because it does not scale for them.
What it costs
Running your own is real work.
You inherit the serving, the GPU bill, the evaluation harness, the security review, the version discipline, the person who understands why throughput fell on Tuesday. A hosted API is somebody else doing all of that, competently, for a price that is usually lower than doing it yourself badly.
So open weights are a trade. You take on operational burden in exchange for control and the ability to specialise.
How to decide
Ask whether your material is unusual enough to need it.
If you are building a general product for general users in English, hosted almost certainly wins by a wide margin. If your value depends on being good at something the frontier models are merely adequate at (your language, your regulated corpus, your niche), then an open model you can tune is the only version of the product that works.
Read the release. Then answer one question honestly: Is our material unusual, or do we just want to feel independent?
One email when we publish. Research, product decisions, and what teams report back.
