Trillion-scale AI models discussed at ADA University

“Instead of training trillion-parameter artificial intelligence models from scratch, university laboratories should focus on adapting smaller models to specialised fields and developing local datasets and benchmarks.”

Huseyin Atakan Varol, Dean of the School of Computing and Artificial Intelligence at Nazarbayev University and founding director of the Institute of Smart Systems and Artificial Intelligence (ISSAI), made the remarks at the 20th IEEE International Conference on Application of Information and Communication Technologies (AICT2026), held at ADA University, TodayPress TV reports.

The speaker said he was honoured to be in Baku, the birthplace of the renowned scientist Lotfi A. Zadeh. He also expressed interest in integration and cooperation with ADA University and other higher education institutions in Azerbaijan.

Varol noted that ISSAI, which has been operating since 2019, was tasked in 2024 with developing national generative AI systems for Kazakhstan. The institute has created the ISSAI Playground platform, which has more than 20,000 active users. Its services include speech and text translation, real-time subtitling, the Olda model for edge devices and the Kupiya tool for protecting personal data.

He also highlighted the institute’s management of Central Asia’s largest code and model repository on Hugging Face, GitHub and ModelScope to support open scientific research in the region. In addition, the institute operates the Irgetas computing cluster based on NVIDIA supercomputers.

Addressing the mathematical foundations of artificial intelligence, Varol stressed that training advanced models requires hundreds of trillions of tokens and substantial computing resources.

The scientist added that smaller models are fundamental to digital sovereignty because they can help protect privacy and operate in local languages.

Jala Rovshan