The Five Layers of ABEJA Platform and Investment in Feature Development for LLMs

「ABEJAはNEDOの公募で採択を得て、データの秘匿性を重視したAI学習手法の開発プロジェクトに参画する。」

「In addition, we continued to work on improving the accuracy of compact LLMs and applying them to physical AI. As a result of these initiatives, total research and development expenses for the interim accounting period amounted to ¥58.697 million.」

「The main initiatives during the fiscal year included the NEDO projects under the Ministry of Economy, Trade and Industry's GENIAC framework: "Model Compression for the Development of Specialized Models" (selected in October 2024 and implemented through April 2025) and "Research and Development on Foundation Models with Long-Context Capabilities and the Construction of AI Agents" (selected in July 2025 and scheduled for implementation through February 2026).」

「During the interim accounting period, we advanced the research and development described mainly in section 2, Business Overview, 2. Analysis of Financial Position, Operating Results and Cash Flows by Management, (3) Operating Results. As a result, total research and development expenses amounted to ¥44.532 million.」

「Against this backdrop, we have allocated resources to LLMs as a focus area, aiming to advance their social implementation by balancing the accuracy and cost required for practical use. This includes building Japanese-language LLMs and conducting research and development on related technologies as an NEDO project under the Ministry of Economy, Trade and Industry's GENIAC framework.」

「For this project, we expect to receive a subsidy of approximately ¥700 million from NEDO, mainly for costs related to the computing resources required to build LLMs. Expenses eligible for the NEDO project subsidy are recorded as accrued research and development expenses, while the corresponding subsidy amount is recorded as a receivable and deducted from research and development expenses. Accordingly, eligible expenses are not expected to affect profit or loss for the fiscal year. During the nine-month period, we mainly conducted research and development on biological AI, deep learning, machine learning, Japanese-language LLMs and related technologies, and research and development activity amounted to ¥24.406 million.」

「Based on the knowledge accumulated to date, we will conduct research and development on Japanese-language LLMs and related technologies, including RAG and agents, with the aim of achieving a substantial improvement in the accuracy and computational cost efficiency essential for the social implementation of LLMs. During the six-month period, we mainly conducted research and development on biological AI, deep learning, machine learning, Japanese-language LLMs and related technologies, and research and development activity amounted to ¥15.319 million.」

「During the three-month period, we mainly conducted research and development on biological AI, deep learning and machine learning, and research and development activity amounted to ¥3.930 million.」

「We have continuously invested in ABEJA Platform since our founding, and investment in its basic functions is almost complete. During the fiscal year, we mainly conducted research and development on large language models, deep learning and machine learning, and total research and development expenses amounted to ¥10.424 million.」

「During the nine-month period, we invested in functional development centered on the five layers comprising ABEJA Platform—the Connect Layer, which inputs the necessary data; the Computing Layer, which performs calculations on the input data; the Security Layer, which processes data to maintain security, including encryption; the Core Layer, which stores technologies for generating software such as AI models; and the BaaS Layer, which prepares general-purpose AI models—as well as LLMs deployed on ABEJA Platform. Research and development activity amounted to ¥6.736 million during the nine-month period.」