Our own AI models — the five FOTOhub hosts and operates itself

Most of the 226 models on FOTOhub come from the 21 providers behind them. Five do not. For IDA Q for images, IDA Voice for speech, the music engine and the image engine we hold the weights, run the inference on GPU servers we operate ourselves, carry the uptime and set the price with no reseller margin in it. The fifth, the Gabriel orchestrator that routes work between models, is our own routing logic and platform knowledge; the small language model it consults to read your sentence is hosted by a cloud provider in the United States.

What "ours" means, precisely

We host the weights, we run the inference, we carry the uptime and we set the price. We are not claiming we trained these models from nothing — that is a different and much larger claim, and a page that blurs the two is not worth reading. Operating them is the part that changes what you pay and where your file goes. Gabriel is the partial case: its routing logic is ours, the language model it consults is not.

The price has no middleman in it

A resold model costs what the provider charges plus whatever margin the reseller needs. On a model we operate there is no second party to pay, which is why these are among the cheapest operations on the platform — and why we can leave one of them free without losing money on every call.

Nobody can deprecate it out from under you

A provider retiring a model breaks whatever was built on it, and the notice period is not yours to set. A model on our own hardware stays until we retire it, a defect in it is ours to fix rather than to report and wait on, and the file being processed does not leave for a third party on the way.

Frequently asked questions

Did you train these models yourselves?
No, and we do not claim to. We host the weights, run the inference, carry the uptime and set the price. Where a model is built on published work, that is what it is.
Can I choose one of them explicitly?
Some are selectable in the studios like any other model; the rest power a feature rather than appearing in a picker. The page marks which is which.
Where does the inference run?
On GPU servers we operate ourselves, for IDA Q, IDA Voice, the music engine and the image engine, so for those the file does not travel to a third-party AI provider to be processed. Not all of those servers are in the EU: voice and music run in the EU, IDA Q runs on cloud GPUs we rent in Switzerland, which is outside the EU, and the segmentation tools of the image engine run on GPUs we rent in the United States. Gabriel is different: no file is involved, but the sentence you type to it is read by a language model that a cloud provider hosts in the United States.
Are they cheaper than the rented models?
Generally yes, because no second party is being paid. The credit cost of every operation is shown before it runs, so you can compare ours against a provider model on the spot.