Starting today, vMira speaks, writes, and bills the way Russian teams need it to. Native Russian, ruble pricing, an on-premise option, and a hosted footprint with primary collection of personal data inside the Russian Federation in compliance with Federal Law 152-FZ.
We have spent two years listening to Russian teams. The question was never whether vMira worked in Russian — Russian was a first-class language from the first model. The question was whether it worked for a company that needed its customer data to live inside the country, its invoices to be ruble-denominated with closing documents on file, and its compliance officers to sleep at night. The answer until now was that the model worked but the surrounding logistics did not. As of today, the logistics work too.
Data residency, no compromises
Per the July 2025 update to Federal Law 152-FZ, primary collection and storage of personal data must happen inside Russia. Subsequent processing — including model inference on already-collected data — can happen abroad if the cross-border transfer is documented and notified to the regulator. Our default architecture for hosted customers reflects that: user requests land first on our Moscow region, where authentication, audit logs, and the primary database live. From there, inference runs either in the same region (default) or, where the customer has documented the cross-border transfer, in a lower-cost foreign region. For customers where any cross-border step is unacceptable, we offer a private deployment option pinned to the Russian Federation, with separate service-level agreements covering availability, incident response, and key management. The on-premise build does not leave the customer's network at all.
Russian, not translated Russian
vMira was trained on Russian press, literature, technical documentation, statutes, and everyday speech — alongside English for transfer, but not as the source language. It understands the register difference between a Moscow office email and a small-town kitchen conversation, and answers in the one you start in. It also handles the major minority languages of the Federation — Tatar, Bashkir, Chuvash, and others — at a level usable for everyday tasks, with steady improvements as we expand the post-training data in those languages. Speech-to-text and text-to-speech in the chat product are tuned to Russian phonology, including stressed-vowel handling and the harder consonant clusters the older open models systematically mispronounced.
“A useful AI is one that sounds like you. And nobody sounds the same from Kaliningrad to Vladivostok.”
Local pricing, local invoicing
All plans are billed in rubles by default, dollars on request. Teams and API contracts include closing documents (УПД) on a monthly cycle, integration with the Russian electronic-document-flow standards (ЭДО), and reconciliation reports formatted for accounting systems used by Russian enterprises. Card, invoice, and SBP (Система быстрых платежей) payments are all supported. Refunds, plan changes, and credit-note issuance are handled in rubles end-to-end without forced conversion. The same plan tiers we offer globally are available — Free, Plus, Pro, Teams, Enterprise — at ruble price points that take into account local card-acquiring fees and the regulatory overhead of operating inside the Federation.
Architecture topology of a single hosted request
Every hosted request lands first on the Moscow region. The edge layer authenticates the user, applies the rate-limit policy, writes the request envelope (user, time, plan, billing entity) to the audit-log database, and assigns a request ID that follows the call end-to-end. From there, two paths diverge by customer policy. Default: the inference call runs in the same region — no cross-border step. Optimised: when the customer has documented the cross-border transfer with the regulator, the inference call is routed to a foreign region with lower marginal cost; the response returns through the same Moscow edge and is logged identically. Either way, the user-facing latency is dominated by the inference itself, not the routing — the edge round-trip is tens of milliseconds, the inference is the rest.
What changed in the 152-FZ July 2025 update
The July 2025 update reaffirmed that primary collection of personal data must occur on infrastructure inside the Russian Federation. It clarified that subsequent processing — including inference on already-collected data — can happen abroad, provided the cross-border transfer is documented in advance and the regulator is notified. Three practical consequences for AI products like ours. First, the database that stores user accounts, conversation history, and uploaded files must be inside Russia, full stop. Second, the inference layer can be either in-region or, with paperwork, abroad — without paperwork, in-region. Third, the audit log of cross-border transfers must be retainable on demand by the regulator. Our default deployment satisfies all three; the optimised cross-border configuration is opt-in and includes the documentation artefacts customers need for their own compliance file.
Roskomnadzor notification: what we file and when
When a customer opts into the cross-border configuration, we file a notification through the Roskomnadzor portal listing the categories of personal data transferred, the destination jurisdictions, the legal basis for the transfer, and the technical measures protecting the data in transit. The filing is straightforward — the categories are pre-templated for our deployment — and customers receive a copy of the filing for their compliance file. Renewal happens annually unless the technical configuration changes; we maintain the calendar and re-file on the customer's behalf. For customers who never opt into cross-border processing, the filing is not required because no data crosses the border.
Ruble billing: УПД, ЭДО, СБП
Three pieces matter in practice. УПД (Универсальный передаточный документ) is the closing document Russian accountants expect at month-end; we issue them automatically on the first business day of the following month, signed with our qualified electronic signature, and route them through the customer's preferred document-flow operator. ЭДО (Электронный документооборот) integration covers the major Russian electronic-document-flow operators, so closing documents land directly inside the customer's accounting system rather than as PDFs in someone's inbox. СБП (Система быстрых платежей) is supported as an inbound payment rail for top-ups; for subscription plans, card-on-file is the default but invoice-on-30-days is available for Teams and Enterprise contracts. Refunds run on the original rail.
Russian voice phonology, calibrated
Speech-to-text in the chat product is calibrated specifically to Russian phonology, including stressed-vowel reduction (the akanye and ikanye patterns), the harder consonant clusters that older open-source models systematically mispronounced, and the prosody differences between formal-register reading speech and informal-register conversation. Text-to-speech goes the other way: stress placement on words with mobile stress (рука/руки/руке) is informed by the morphological analysis the model already performs for text. Latency stays under two hundred milliseconds for the typical phrase on a 5G connection. Minority-language speech is on the roadmap with Tatar and Bashkir landing first; today, those languages are text-only at usable quality.
Minority languages of the Federation
Tatar, Bashkir, Chuvash, and several other languages of the Federation are usable today for everyday tasks: drafting, summarisation, translation, simple reasoning. They are not at the same quality level as Russian — the post-training corpus is materially smaller — but they are real, not a fall-through to translated Russian. Our roadmap adds steady improvements every quarter, paired with public benchmark scores on each language so customers can track progress without taking our word for it. If you have a vertical use case in any of these languages, we want to hear from you; targeted post-training data is the largest lever we have on quality, and we prioritise the use cases that customers tell us are blocked.