What is required
Typical security team requirements: data and the model inside the perimeter or in a certified environment, and a Russian model or a model with open weights whose makeup can be checked. Since 1 September 2026, Russia’s AI law 243-FZ has been in force, with the statuses of sovereign and national models and criteria from 1 March 2027. A working group discussed a draft government decree on mandatory use of sovereign and national models in public administration, tax and customs, health care, education and CII at the end of August 2026; Russia’s Ministry of Digital Development says the list of sectors has not been set.
Who has already done it
In February 2026, a telecom operator announced a tender for an AI developer assistant in a closed data center: 15 users, a response time of up to 400 ms, refactoring accuracy of 85% or more, the MCP protocol. An airport built an AI plugin for secure development into a closed environment, with more than 90% accuracy in checking findings and up to 85% in fixing them. The Ministry of Justice is training a local model for a lawyer avatar with a base of federal laws. Moscow’s health system built an AI assistant on a local model into the data marts of its hospitals. A developer and a bank set up an AI controller for project finance in an isolated environment and cut documentation analysis by 80%. The operator of an insurance system deployed a local model for requests, support and code review and handled growing load without adding staff.
What to install
For a department of 10–20 users, by our calculation: a server with two 96 GB cards, about $30,000 as of 29 September 2026, a 70-billion-parameter model in 8 bits, while a 235-billion model in 4 bits only just fits for 10 users with an 8,000-token context; vLLM; hosting in your own data center. For hundreds of users: a server with four H100s from $120,000, or several of them. For a pilot: one workstation for about $13,000, or a rented card in a certified Russian cloud if the policy allows. An engineer is a must: by our estimate, 10–20% of a position per server and a full position per cluster.
What to write in the tender specification
The number of users and parallel requests. Latency to the first token and generation speed. Accuracy on your set of 50 requests, with a threshold. A list of data that does not leave the environment, and the anonymization method for the rest. The option to replace the model without rewriting the agents. Monitoring, redundancy and a plan for a card failure. Request logging for audit. The telecom operator’s tender has similar points: the number of users, response time, accuracy and model choice.
What to expect from us
A pilot on your hardware or in a certified cloud in 14 days from $1,499 with a test on 50 requests, a hardware specification from the calculators, agents through one model interface, and handover to your engineer. Your lawyer gives the legal opinion under 152-FZ and 243-FZ; we provide the architecture and documentation.