Most "AI infrastructure" quietly ties a company to two things at once: one model's weights, and whatever harness happens to be wired to them. Change the model and the harness breaks. Outgrow the harness and you are stuck with the model it was built around. lumina exists to cut both leashes at once — any model, on a harness fast enough that the choice of model stops being a compromise.
That is not a dockerized, stripped-down Hermes sitting in a jail cell. Both Hermes and the harness were rebuilt around a decision engine. Before a language model ever sees a request, and alongside it while it streams, Jev — or its Laya-compatible sibling — runs a typed decision: a real, scored choice among concrete options, answered in tens of milliseconds, not a paragraph of prose to be parsed and hoped for. The fast lane hedges providers instead of calling them one at a time, so whichever one answers first wins. Speculative intents read the assistant's own delta stream and start acting mid-response, before the sentence finishes. And policy sits outside the model entirely: a request the policy engine has not cleared never reaches the model, so nothing has to be talked out of a bad idea after the fact.
The consequence is what makes the container cheap to run. Because no model weights ever run locally — only the call to the language model leaves the box, out to the cloud — a company's lumina container does not need a VPS's worth of GPU and memory to exist. A micro VM is enough to hold Hermes, the harness and the decision engine; the model itself stays exactly where it already lived.
And the deal gets better the longer it runs: after 3 months of continuous subscription, every rate on the account drops 10 percent, automatically, for as long as the subscription continues.