Tsuzumi 7B — это компактная и эффективная языковая модель от NTT, вдохновленная эстетикой японского барабана. Этот ИИ доказывает, что сеть из небольших специализированных нейросетей может решать сложные задачи эффективнее, чем одна громоздкая модель.
tsuzumi is a large-scale language model created by NTT Laboratories. The name is inspired by the traditional Japanese drum “鼓” and the model reflects the instrument’s compact and efficient design. Our vision for the future involves tackling societal challenges through the collaborative intelligence of a network of smaller, specialized LLMs like tsuzumi. In this presentation, Kyosuke Nishida, Senior Distinguished Researcher in the NTT Human Informatics Laboratories, demonstrates the tsuzumi-7B model, which was developed from scratch and features 7 billion parameters and over one trillion Japanese and English tokens. A vision-and-language model using tsuzumi for visual document understanding is also showcased.