Open-weight by default, frontier models by choice
Match model access to quality and data requirements, with external providers enabled explicitly.
Editorial analysis and illustrative technical scenarios. They do not define the features or integrations included in a UTOMAI subscription. See Private AI On-prem pre-launch information.
AI-generated editorial illustration. Not a product photograph or a customer case study.
A private AI platform should let the company discuss model quality and data handling together. An impressive answer is not sufficient if the processing route conflicts with the agreed boundary. Equally, a local model still needs to be useful on the team’s actual tasks.
Start from the planned open-weight catalogue
UTOMAI private AI uses selected open-weight model families on dedicated cloud or on-premise infrastructure. The configuration confirms processing locations, compatible hardware, runtime and licence conditions. Open weights do not make every model free to redistribute or suitable for every machine.
Request external frontier models deliberately
An external model from a European or non-European provider can be assessed when a task needs a different capability or quality level. Confirm the endpoint, allowed information, terms and costs. The workspace can remain the common working environment while each permitted route has a distinct data boundary.
Do not hide a provider change in failover
Continuity policies should use approved alternatives and respect the agreed geography and data restrictions. A failure in a local or European route is not permission to send the request to any available endpoint. When an acceptable alternative is unavailable, queueing, retrying or reporting a temporary limitation may be the correct response.
Compare permitted models on representative questions, source use and correction effort. Record the reason for the choice and review it when the model or requirements change. The goal is useful AI within a known boundary, rather than a default preference for the largest model.
