Where an open model is just as good
Short chains of actions: in 27 runs by the IT unit of a store chain, a local model with 9 billion parameters was as good as cloud models on chains of 2–3 steps. Extracting fields from documents, sorting requests, answers from a knowledge base with search, email drafts: in our estimate, models of 8–32 billion parameters handle these; test on your own data. Russian language: the open 32B model of a large Russian bank and the open 8B model of a large Russian technology company are built for Russian. Tasks with a fixed answer format where the model was fine-tuned on your own data: the legal department of an IT holding trained a model on its own contracts for more than a year and cut contract approval from two weeks to three days.
Where it loses
Long chains: on 15 steps, no model gave a fully verifiable result, but a large 235-billion cloud model gave partial success in 2 runs of 3, and the 9-billion local model in 1 of 3, with 2 false reports of completion. Complex code and architecture decisions, legal analysis of long contracts, reasoning with dozens of conditions: for these, people pay $51 per million output tokens for Fable 5.1 and GPT-6 Astra. A vision model and speech at the same time: locally this means two models and memory for both. Speed on a long context on one computer: in a 2025 measurement, a 671-billion model on a Mac Studio M3 Ultra with 512 GB reads 8,000 tokens in almost 15 minutes.
Price per token
At 60 million tokens a month, cloud models cost from $12 on GPT-6 Luna to $1,200 on flagships, and $900 on a large Russian cloud model. An open model on a workstation with an RTX 5090 costs about $800 a month including an engineer, at any volume up to the card’s limit: $5 per million tokens on an 8-billion model in one stream at 50% load. A 70-billion model does not fit on one card; on a workstation with two RTX 5090 cards at about $1,100 a month, it costs $46 per million. The open model wins on money at volume with mid-size models and loses at low volume and against cheap cloud models.
Data and the law
An open model on your own hardware: data does not leave the perimeter, you need no processing contract with the model vendor, and the architecture ensures data localization, while your other duties as an operator under 152-FZ, Russia’s personal data law, remain. Foreign flagships: Russian addresses are not in their supported regions, payment goes through intermediaries, and most of them have no processing contract. Russian clouds: a contract with a Russian legal entity and data in Russia, but the setup belongs to someone else. Under 243-FZ, the developer of both a sovereign and a national model is a Russian legal entity, and answers and data are processed in data centers in Russia owned by Russian legal entities; a national model may include components under an open license, including other people’s open models. The status goes to the model and its developer, not to the company that runs the model on its own hardware; these rules come into force on 1 March 2027.
How to choose
Split the task into steps and count how long the chain is: on chains of up to 3 steps, the open model in the test was as good as cloud models; for long chains, test the cloud or a hybrid. Take 50 of your own requests and run them on three models: an open 8–9 billion model, an open 30–70 billion model and a mid-range cloud model. Count not the price per token but the price of a solved task: a model that is wrong in 10% of cases costs as much as a person’s time to fix those 10%.