companies

Preferred Networks

Tokyo-based deep learning company building the PLaMo family of Japanese-first foundation models alongside custom AI chips and supercomputing hardware.

Preferred Networks was spun out of Preferred Infrastructure in March 2014 by University of Tokyo computer science alumni Toru Nishikawa and Daisuke Okanohara, who wanted to apply deep learning to industrial problems in robotics, manufacturing, and healthcare. Based in Tokyo's Otemachi district, the company built its early reputation on the Chainer deep learning framework before Chainer development wound down in favor of PyTorch. PFN has raised roughly $300 million in total funding from investors including Toyota, Fanuc, NTT, Nvidia, and Hitachi, and it builds its own AI accelerator hardware — the MN-Core chip family and MN-3 supercomputer — alongside its models, an unusual degree of vertical integration for a company its size. Since 2023 PFN has focused heavily on PLaMo, a large language model line trained from scratch on Japanese-heavy data, which it says outperforms GPT-4 on several Japanese benchmarks including national medical licensing exam questions; PLaMo has since been adopted by Japan's 'Gennai' government AI project.

Founded

2014

Headquarters

Tokyo, Japan

Sector

lab