How it works
A router with one interface stands between the agents and the models, so an agent doesn’t know which model will answer. The routing rules: by data type, personal data and code always stay local; by complexity, chains of 10 or more steps go to the cloud; by time, night batch work runs on the local model; by failure, if one model is down, requests go to another. Before anything goes to the cloud, the router anonymizes it: it replaces names, numbers and amounts with labels and puts the real values back into the answer.
What it gives in cost
An example for 60 million tokens a month: 80% simple requests on a local model on a workstation with an RTX 5090, at $800 a month with an engineer, and 20% complex ones on Sonnet 5.5 at $49. The total is $800, against $1,200 on Fable 5.1 for the whole volume and $250 on Sonnet 5.5 for everything. So the hybrid is cheaper than the flagship and more expensive than a mid-range cloud model. Companies choose it for data control and fault tolerance, and by our calculation it wins on cost as volume grows.
What it gives in legal terms
152-FZ, Russia’s personal data law: personal data stays inside your perimeter, and anonymized data sent to the cloud is still personal if the labels can be linked to a person. So it needs a Russian provider with a data processing agreement, and a lawyer gives the final opinion. 243-FZ: the law gives sovereign and national model status to a model of its developer, a Russian legal entity, when answers and data storage happen in data centers in Russia. For the local part, this is a reason to choose a model that will get this status; the criteria take effect on 1 March 2027. It matters if a draft decree is adopted that would require sovereign models in some sectors when they cost no more than 1.5 times the national ones; the draft was reported on 8 September 2026.
What it looks like in cases
A telecom operator required in a tender that source code never leave its data center, with code completion responses within 400 ms. An airport built an AI plugin that checks and fixes code in a closed network, with more than 90% precision on checked findings. A retail chain compared a local 9-billion-parameter model with cloud models in the same setup: on a two-step task, all three models succeeded in every run. A property developer put a Russian model in an isolated network to check project finance terms and cut document analysis time by 80%.
Where to start
Build agents through one model interface, even if everything is in the cloud today. Label your data: what must never leave. Put a local model on one class of tasks and measure quality on 50 requests. Turn on routing and anonymization. We build this in a pilot in 14 days from $1,499 and in an AI department in 4–6 weeks from $4,999.