In Russia, Cotype 3 AI Models Bring Multi-Step AI Agents Closer to Enterprise Reality
MWS AI, the artificial intelligence division of MTS Web Services (MWS), has introduced CotypePro 3 and CotypeLight 3, its third-generation large language models. Both models can process text and images simultaneously and are designed to power AI agents capable of handling complex, multi-step workflows. According to industry experts, they could help businesses respond to customer inquiries faster, accelerate document preparation, and automate tasks in human resources and banking.
According to the developer, CotypePro 3 and CotypeLight 3 feature 27 billion and 9 billion parameters, respectively. They were built to support enterprise-grade AI agents capable of completing multi-step tasks autonomously without human intervention. A key advantage is deployment flexibility: organizations can run the models entirely within their own secure infrastructure, including servers powered by the Russian AstraLinux operating system, and then fine-tune them using proprietary corporate data.
Built to Remember Large Documents and Keep Learning
The models can operate independently or as part of the MWS AI AgentsPlatform. Both support a context window of 262,000 tokens, allowing them to keep an entire package of contracts for a major transaction or a company's annual report with all supporting appendices in working memory at the same time. They can run on Nvidia A100 GPUs, making AI agents based on CotypePro 3 and CotypeLight 3 accessible to a broad range of organizations. MWS AI trains both models using the computing resources of MWS Cloud. During testing, the Cotype family also demonstrated full compatibility with every component of Russian hardware-software platforms.
Evaluated Across 178 Business Scenarios
To measure the agent capabilities of the new models, MWS AI created its own benchmark consisting of 178 business scenarios. Each task assigns the model one of five workplace roles and evaluates how effectively the AI agent uses available tools, follows business rules, and reaches the required outcome. Depending on the scenario, the agent serves as a customer support specialist handling subscriber questions about services, subscriptions, and billing; a tariff consultant recommending or changing service plans based on customer needs; an HR specialist calculating vacation and sick leave benefits while processing personnel documents and company analytics; or a tax risk inspector investigating schemes that shift business activity between companies and determining when tax audits should be initiated.
The evaluation measures two key indicators: the percentage of scenarios in which the model successfully completes its assigned task and a reproducibility score showing how consistently it reaches the same result across repeated runs. CotypePro 3 achieved scores of 92.2% and 79.1%, respectively, while CotypeLight 3 reached 88.6% and 73.9%. Both significantly outperformed earlier generations of the Cotype models. In the independent MERA benchmark, which compares language model performance, the 27-billion-parameter model ranked third despite competing against systems several times larger. In the open MWS AI Vision Bench, which evaluates work with Russian-language corporate documents, it achieved 74% accuracy, outperforming foreign models. The benchmark assessed text extraction from images, document structure reconstruction, page layout recognition, data extraction, and question answering across documents written in Russian, English, and Chinese.
Enterprise AI for Banks and Government Services
As organizations adopt the Cotype family of models, customers could benefit from faster response times, quicker service recommendations, and more efficient document preparation. The primary users are expected to include banks, telecommunications providers, industrial enterprises, and government agencies.
The ability to deploy the models entirely within a customer's own infrastructure could reduce reliance on foreign software and cloud services across both business and government. The new models also have export potential: they support Russian, English, and Chinese, while allowing organizations to keep confidential information inside their own infrastructure instead of transferring sensitive data to external cloud platforms. In the future, likely export markets include CIS countries, Central Asia, and selected BRICS members, where demand exists for Russian-language AI solutions or locally hosted AI systems.
