MTS AI unveils Cotype Pro 2, a new version of its business-focused large language model Cotype. The new model performs 40% faster compared to the previous iteration and produces 50% more accurate results when processing long texts of up to 128 thousand tokens (this is equal to about 900 thousand characters with spaces). It is capable of running over 100 scenarios that Russian corporate users may need. The product can be deployed on-premise with no connection to external servers required, to ensure data security. The model was among the top five Russian LLMs in the MERA benchmark by AI Alliance.
In 2025–2026, MTS AI plans to use the Cotype Pro 2 model to launch its new line of AI solutions for the following seven industries: public sector, banks, manufacturing, retail, telecom, medicine and IT. In the first stage, the company will focus on AI assistants, and next — AI agents, highly autonomous services intended to cover both general (HR, customer support, document flow, etc.) and industry-specific tasks.
“Аrtificial intelligence technologies are shifting to new architectures possessing completely different capabilities. We need our new generation models to be capable of not just processing information but also understanding the physical world, interacting with it on a deeper level. For instance, they can analyze data from various sources, such as video surveillance systems or production sensors to forecast possible failures or optimize processes. Companies wishing to eliminate downtime or improve performance will be particularly interested in this feature. AI also requires ‘long-term memory.’ It needs capabilities to memorize user preferences and communication history to provide more customized responses. Contemporary models must also possess more complex planning features to efficiently solve multistep tasks, such as project management or teamwork coordination,” noted Dmitry Markov, Executive Director, MTS AI.
“The key here is to implement a robust reasoning ability so that our models can analyze, draw conclusions and make informed decisions among uncertainties, for example, when evaluating risks in financial analytics or finding optimal approaches in scientific research. AI agents must be able not to simply follow instructions and process large amounts of information but also give structured answers that other systems can understand. We developed a dedicated benchmark to assess the quality of these answers. It features tasks of various complexity and context size. Our test showed that Cotype Pro 2 solved 94% of the tasks, while GPT-4o, 93%,” explained Sergei Ponomarenko, LLM Product Director, MTS AI.
Prior to release, MTS AI spent three months testing the Cotype Pro 2 model in real conditions: as an automatic support request sorting tool in a bank, and as an event description creation feature in MTS Live. These tests also confirmed that the model can be integrated with RAG (Retrieval-Augmented Generation) tools to retrieve information from databases with text generation.
Cotype Pro 2 greatly surpasses the previous model in a number of tasks. As such, its idea generation performance increased from 44% to 57%. Its information retrieval accuracy rose from 81% to 86%, demonstrating a better performance in key data selection. In terms of general discussions, the model showed an improvement from 55% to 64% and a smoother and more meaningful interaction. At the same time, classification performance remained at a stable high level of 87%. Improvements in summarizing (from 79% to 85%) and creative writing (from 61% to 65%) also highlight better skills of text processing and making substantial creative materials.
Apart from basic functions (summarization, classification, reasoning, mathematics and logics), Cotype Pro 2 comes with a ‘role-playing’ feature. It is capable of having a conversation acting a specific character or specialist by adapting the communication style the relevant profession has. It improves interaction in customer support, personnel training and other areas.
