Zehnlab.ai, the developer of Soro, a large language model for the Tajik language, has been named one of the five winners of NVIDIA’s Open Models AI Codefest. The competition focused on developing artificial intelligence solutions for the public sector using open models, according to T-Systems, a subsidiary of Deutsche Telekom.
NVIDIA organized the competition with support from Deutsche Telekom and T-Systems. The organizers received 350 applications and selected 14 teams from 13 countries across Europe, the Middle East and Africa. Participants spent five weeks developing AI solutions using open models and NVIDIA Nemotron libraries.
Following the final Demo Day on October 7, the jury selected five winners: Pleias of France, zehnlab.ai, ABA Technology of Morocco, Disseqt AI of Ireland and Algoritmica GmbH of Germany.
The winning teams presented their projects at the NVIDIA GTC Berlin Public Sector Forum on October 20. NVIDIA also named zehnlab.ai among the top five Codefest teams in a blog post about open models for Europe’s public sector.
Agentic Soro to serve more than 4,000 schools
Zehnlab.ai entered the competition with Agentic Soro, a project designed to turn Soro into an AI agent capable of using digital tools and searching for information online in response to queries in Tajik.
According to T-Systems, the team developed the system using open NVIDIA Nemotron datasets. Agentic Soro is intended to become the next version of Soro, which is being rolled out to more than 4,000 schools.
The system is designed to provide answers with source references, allowing users to check the information themselves.
During the competition, T-Systems provided participants with access to NVIDIA RTX PRO 6000 graphics processors through its Industrial AI Cloud infrastructure in Munich, operated by Deutsche Telekom.
Soro expands across Tajikistan’s schools
The number of schools using Soro initially grew from 25 to 100 before reaching 4,000 in September 2026, in partnership with Tajikistan’s Ministry of Education and Science.
The platform helps students study educational materials in Tajik. The developers also plan to introduce tools that will enable teachers to prepare lessons and tests.
By 2027–2028, the team aims to make Soro available to more than two million students.
Developing AI for underrepresented languages
Soro is developed by an international team comprising specialists from Tajikistan and experts from Oxford and Cambridge universities. Zehnlab.ai has offices in the United Arab Emirates and Tajikistan, while the model’s computing infrastructure is provided by the Darya.ai cluster in Tajikistan.
The company is also adapting its approach to other languages for which limited training data is available. T-Systems has described Agentic Soro as a potential model for developing AI solutions for such languages.
Zehnlab.ai develops its technologies through Area AI, an artificial intelligence hub officially launched in Dushanbe in January 2026. Its work includes language models, AI agents and synthetic data technologies.
The company also focuses on AI safety. Its Soro Guard model is designed to detect unsafe content in queries and responses in Tajik, Russian and English.
Zehnlab.ai is a member of the NVIDIA Inception program and received the AICA 2025 award in the Best AI Development category. It was also a finalist in Saudi Arabia’s Misk Accelerator and Germany’s Constructor Start program.