Through this project, we will develop a sovereign multimodal foundation model that will underpin the development of AI robots and Physical AI. In doing so, we aim to strengthen Japan’s generative AI development capabilities, enhance the competitiveness of manufacturing and other industries, and contribute to green transformation. We will pursue AI foundation model development in tandem with pioneering research into theories, component technologies, and architectures, conducted in collaboration with research institutions in Japan and overseas. We will progressively develop and release foundation models equipped with core capabilities such as advanced Japanese-language understanding, logical reasoning, and instruction following; multimodal foundation models capable of understanding images, video, and audio; and world foundation models capable of perceiving physical space.

This project will be undertaken by a consortium comprising Noetra Inc. and the National Institute of Advanced Industrial Science and Technology (AIST). Within the consortium, Noetra Inc. will lead the Development Track, focusing on the development of multimodal foundation models, while AIST will lead the Exploration Track, undertaking pioneering research and development, with the two tracks closely coordinated as part of an integrated initiative. Under the Development Track, Noetra Inc. will leverage its shareholder base, which includes a large number of companies expected to use the foundation models, to advance development based on real-world user needs. Under the Exploration Track, AIST will draw on its network of leading researchers and research institutions in Japan and overseas to conduct pioneering research into theories, component technologies, architectures, and related areas through joint research with a broad range of domestic and international research institutions.

The models developed through this project are expected to be released as project outputs in each fiscal year. They are intended to be made available as weight-available models, with access to the trained weights limited to entities that meet specified eligibility requirements. This approach reflects the view that, because the project’s outputs are being developed as part of a government-funded initiative and will incorporate Japan-specific data, they should be provided under appropriate controls rather than released without restriction.