Preferred Networks, Inc. is a Japanese artificial intelligence company founded in 2014 and based in Tokyo. Its approach rests on a conviction: advances in deep learning must translate into practical applications, particularly in the physical world. With an online presence at preferred.jp, it works on both models and the infrastructure needed to train and run them.
Software roots, industrial foundations
Preferred Networks grew out of Preferred Infrastructure, a company specializing in information processing technologies. Co-founded by Toru Nishikawa and Daisuke Okanohara, it expanded by building ties with major Japanese groups, including Toyota and FANUC. These collaborations placed it at the intersection of artificial intelligence research, robotics and production constraints.
The company also gained recognition in the scientific community with Chainer, an open-source deep learning framework launched in 2015. This tool made it possible to build computational graphs dynamically, an approach well suited to experimentation. Preferred Networks later shifted its development efforts to PyTorch while continuing its research and publishing activities.
From chips to models, an integrated approach
Unlike companies focused on a single software layer, Preferred Networks also develops hardware. Its MN-Core processor family, designed with Kobe University, aims to accelerate deep learning computations, notably by improving their energy efficiency. This expertise in computing supports its research into models and applications, although it entails investments and skills distinct from those of a conventional software vendor.
In materials science, the company developed the Matlantis platform with ENEOS, commercially available since 2021. It uses machine learning to perform atomic-scale simulations and explore the properties of materials. The aim is to help research teams select promising avenues before conducting extensive experimental testing, without removing the need for physical validation.
Preferred Networks is also active in large language models with PLaMo, a family of models whose first publicly released work dates back to 2023. The Japanese language is a major focus of this initiative. More broadly, its activities span industrial, robotic and scientific applications, with an emphasis on the transition from prototype to usable system.
What next?
The next stage will depend on the company’s ability to turn this technological depth into recurring use. In industry, reliability, integration with equipment and operating costs matter as much as model performance. In language models, Preferred Networks will also have to contend with strong international competition. Its Japanese roots, industrial partners and computing expertise are assets, provided it can demonstrate their value in specific tasks. The challenge is less about covering every AI use case than about making its technologies relevant in demanding environments.