How MTS AI saved more than a billion rubles in investment through model training in the MWS cloud

2025: Moving AI Models to the Learning and Inference Cloud

MTS Web Services on May 14, 2025 announced that MTS AI has completely transferred to the cloud the process of training and inference of artificial intelligence models, as well as large language models. This allowed the company to accelerate the launch of products and save more than a billion rubles in investments in its own infrastructure.

MTS AI is one of the competence centers in AI. Russia The company develops products and solutions based on, generative AI natural language processing technology and. computer vision Among the company's services: AI service - speech analytics WordPluse, synthesis platform speech recognition and based on neural networks and methods - machine learning Audiogram, code generation and auto-completion service for optimizing the development process - as well as Kodify voice text assistants for automating business communication with customers.

All of the company's services are based on ML models that are trained on thousands of terabytes of data. To speed up the process of training models, the company uses GPU. Self-procurement of equipment requires serious investment in infrastructure, so MTS AI decided to move the learning process to the cloud. An ML platform is deployed on the MWS virtual infrastructure, in which all key models of the company are trained. After completing the training, the models continue to be infected in the cloud, which allows not only to bring the service to the market, but also to support its work.

To solve more complex problems, MTS AI offers its clients services based on large language models. Among them are LLM for working with texts, as well as searching and analyzing information - Cotype and its lightweight version - Cotype Nano. As well as an on-premium service for generating and completing code to optimize the development process - Kodify. LLM training is more time-consuming than working with conventional ML models. Training and further training of LLMs on consumer GPUs can take considerable time. To speed up this process, MTS AI uses the power of the MWS supercomputer - MTS GROM. It allows the company to complete its own LLMs tens of times faster.

The work of MTS AI with models in the cloud is divided into three stages. On the first - there is a departure and pre-training of models. This allows you to prepare models for intensive training and reduce the cost of it. This mainly uses a virtual infrastructure with GPU accelerators. In the second stage, the models undergo deep training. It takes place on a supercomputer, some of the company's neural networks are further trained on the VI c GPU. After deep training, the models are ready to work in the company's products and services. In the third stage, the models switch to inference, which also occurs in the MWS cloud.

A large number of different equipment is required to work with artificial intelligence models. The more models you implement, the more diverse it is. For simpler models, weaker cards are needed, for more powerful ones, more productive ones, for LLM training, a supercomputer, for an inference, another type of GPU can be used. When working with on-prem models, all these capacities would have to be purchased independently or work with all models on one form of GPU, which would lead to a significant increase in investment costs or the irrational use of infrastructure. Moving to the cloud allows us to use exactly the amount of computing resources that is necessary and not spend huge amounts on the purchase of servers. Only in order to purchase equipment for LLM training, we would have to invest more than a billion rubles in infrastructure. And thanks to the use of MWS capacities, we can redistribute investments to new projects. In addition, the cloud consumption model allows us to accelerate the launch of new products, as there is no need to wait for the delivery of equipment, "said Sergey Ponomarenko, director of LLM products at MTS AI.