A private, on-premise language model runs on your own server or in a hosted environment inside Russia instead of someone else's cloud, so prompts, contracts and customer records never leave your control. In 2026 you no longer need to train a model for this: Sber and Yandex have released open weights that companies can run in-house and use commercially. The real question is not which model to pick, but which data you are willing to send outside and how much hardware you are prepared to run.
What changed: open Russian models
Sber publishes the GigaChat family on Hugging Face under the MIT licence. The compact GigaChat3-10B-A1.8B has 10 billion parameters, of which 1.8 billion are active per token. In July 2026 Sber released GigaChat 3.5 Ultra with 432 billion parameters (28 billion active). Yandex has opened YandexGPT 5 Lite 8B under its own licence, which allows commercial use on stated terms, so read them before you deploy.
Model size drives the budget. A 10-billion-parameter model in fp8 needs roughly 10 GB of memory and fits on a single server GPU. A 432-billion-parameter model in the same format needs hundreds of gigabytes of GPU memory, which means a cluster. Most everyday business tasks are well served by the compact models.
When you need a local model and when you don't
Part 5 of Article 18 of Federal Law No. 152-FZ requires personal data of Russian citizens to be recorded and stored in databases located in Russia when it is collected. Sending that data to an external AI service counts as outsourced processing under Part 3 of Article 6 of the same law: you need a contract with the provider and a clear picture of where the processing physically happens. A local model takes that question off the table because the data stays with you.
- worth it: reviewing contracts and client correspondence, an internal knowledge base, HR documents, anything your security team won't let out of the building;
- optional: website copy, anonymised analytics, email drafts with no personal data.
Rule of thumb: split your tasks into "fine to send out" and "must stay inside". A local model is for the second group; the first is cheaper to handle through a cloud API.
How to deploy a local model, step by step
- Pick one or two tasks and list the data they touch.
- Choose a model for the task and test it on your own documents, not on a demo.
- Set up a server, either your own or rented in a Russian data centre, reachable only from your network.
- Connect the model to the system people already use. In sales that is the CRM: the agent reads the deal card and writes its answer back there, as in on-premise Bitrix24 on your own server.
- Name an owner who updates the model, tracks answer quality and manages access.
The classic mistake is starting with the biggest model available. It costs more in hardware, and on a narrow job like sorting inbound requests a compact model with good instructions often does just as well. The second mistake is running the model in isolation from your processes, so staff keep copying text back and forth by hand.
What it means for your budget
A local model has three cost lines: hardware or a rented GPU server, setup work, and ongoing maintenance. The open GigaChat licence costs nothing, but a GPU server is far more expensive than a regular one, so payback has to be worked out on a specific task: how many staff hours it frees up each month. Deploying one AI agent on your process starts at 97,500 RUB with us, and extra work is billed at 3,900 RUB an hour; see our AI agents page and pricing. If the task involves no personal data, a cloud option is usually cheaper to start with, and companies move on-premise once request volume and security requirements justify it.