ABEJA's Innovative Leap in AI Tech
Introduction to the GENIAC Project
ABEJA, a Tokyo-based company driven by its philosophy of creating a "rich world through implementation," has recently concluded the third phase of the GENIAC (Generative AI Accelerator Challenge) project. Co-supported by the Ministry of Economy, Trade and Industry and NEDO, this initiative aims to bolster the development capabilities of generative AI in Japan.
The third phase has seen ABEJA leverage insights gained from the earlier phases to create large language models (LLMs) and AI agents that can be effectively used in mission-critical operations, environments where errors cannot be afforded.
Research and Development with Industry Leaders
As part of this project, ABEJA collaborated with leading firms, SOMPO Holdings and Sompo Japan Insurance, to explore the potential applications of their developed LLMs in the field of insurance. This partnership aims to test and validate the model's effectiveness in real-world scenarios.
The Digital Platform Business of ABEJA
Since its establishment in 2012, ABEJA has committed to advancing deep learning technologies and has been actively involved in groundbreaking research in various fields, including LLM and quantum computing. The ABEJA Platform has been designed to aid in the introduction of AI into mission-critical operations, enabling organizations to enhance their operational efficiency.
Addressing Challenges of Mission-Critical Business
Mission-critical operations are sensitive by nature and require precise, expert handling. Companies often hesitate to implement LLMs due to security concerns regarding data management. With a growing emphasis on protecting sensitive information, ABEJA seeks to explore and develop LLMs suitable for secure environments.
The belief is that applying LLMs to mission-critical operations can transform industry structures. During the third phase of GENIAC, ABEJA thus focused on research and development of secure LLMs and surrounding technologies applicable to such operations.
Achievements of the Third Phase
The third phase yielded significant outcomes outlined below:
1.
Compact LLM and AI Agent Development
ABEJA developed a compact LLM known as the "ABEJA Qwen3-14B Agentic Model" based on Alibaba's Qwen series. This model functions within a company's on-premise environment, removing the need to send sensitive data externally, thereby maintaining data sovereignty and security. The model possesses advanced Japanese language capabilities and excels in processing long contexts, outperforming competitors such as GPT-4o and Qwen3-14B.
2.
Proving Operational Integration
In collaboration with SOMPO Holdings and Sompo Japan Insurance, ABEJA applied the LLM to specific tasks in the insurance sector. This exploration demonstrated that even in a highly specialized field demanding security, independent reasoning and outputs could be achieved, validating its potential as a practical tool in real-world applications.
3.
Knowledge Sharing Initiatives
ABEJA actively participated in events and seminars held by NEDO, sharing insights and knowledge gained from the third phase's research and development. Comprehensive details about the phase's outcomes and methodologies are accessible through the provided links.
Hugging Face:
Blog:
Future Directions
ABEJA remains committed to pushing the implementation of LLMs in mission-critical operations forward, continuing its efforts to fulfill its corporate mission of creating a "rich world through implementation."
The insights gained from this project are a result of NEDO's funded initiative.
For further information about ABEJA, visit:
ABEJA Official Website.
Company Background
Founded in September 2012, ABEJA operates out of Minato-ku, Tokyo. Led by CEO Yosuke Okada, the firm focuses on providing AI support for mission-critical operations through the ABEJA Platform, a robust system designed to enhance the collaboration between humans and AI.
Aiming to transform industrial structures with technological power, ABEJA has earned the trust of various enterprises that have integrated their solutions.