Small Open Models Are Eating the Middle, and That's Where I Build
GLM-5.3 and the decider class show the action isn't at the frontier. It's in the models light enough to own and tune.
The frontier gets the headlines, but the build-worthy move this month is the middle. Z.ai's GLM-5.3, a 744B open-weight model, drew Anthropic's 'most cyber-capable open weight' label, and the cheap decider models show you don't need a giant for most steps. The middle got good.
For someone shipping product, that's the useful zone. A model I can self-host or fine-tune beats a black-box API when the task is stable and the data is mine. The frontier is for the hard 5%; the middle handles the other 95% at a fraction of the cost and with none of the vendor risk.
Open weight also means the cap on a task is my engineering, not a rate limit or a policy change. When a lab gates its best model (see Argon), the open middle is the fallback that keeps me building instead of waiting.
I'm moving more of my stack to models I control. The frontier is a service; the middle is a tool. For most of what I ship, I'd rather own the tool.