Product · November 5, 2026
Separator types a request, then moves it in milliseconds
The first public description of Separator: a type on every request, a model chosen for that type, and a split when one request is really many.
Separator is the routing layer VeloxitySI integrates in front of a catalog of models. A request arrives as text, audio, or a media prompt. Separator gives it a type. The type is what the next hop is allowed to be.
The decision is the product. It is designed to land in milliseconds, and it names the model that should do the work. A question about the sky goes to a reasoning model. A patch goes to a code model. A still goes to a vision model. The caller can read the type, the model, and the time.
When the request is doing more than one job, Separator does not pick a compromise model. It splits the request into small, typed calls and sends each call to the model that is good at that part. The response is the list of those calls, not a blend.
Model names in the catalog, reason-1, code-1, write-1, search-1, vision-1, speech-1, and ops-1, are lanes. An allow-list on the request decides which lanes may be used. A lane that is absent cannot be chosen.