FIXER Accelerates Social Implementation of Japanese-Optimized LLM "GPT-OSS Swallow" Through Joint Research with Tokyo Institute of Technology

FIXER Inc. (Headquarters: Minato-ku, Tokyo; Representative Director and President: Seiichi Matsuoka; hereinafter "FIXER") is pleased to announce the signing of a joint research agreement with the Okazaki Laboratory at the Institute of Innovative Research, Tokyo Institute of Technology (Location: Meguro-ku, Tokyo; President: Naoto Ohtake; hereinafter "Tokyo Tech") to further advance their collaborative efforts in the social implementation of large language models (LLM) that began in 2024 (at the then-Tokyo Institute of Technology, now Tokyo Tech). This new agreement focuses on accelerating the social implementation and sophistication of "GPT-OSS Swallow," an LLM optimized for Japanese language.

In this joint research, FIXER will optimally implement "GPT-OSS Swallow," which is available for commercial use, into its on-premise generative AI solution "Sovereign GaiXer" and verify its effectiveness. This will accelerate the realization of practical and high-performance generative AI utilization in fields requiring extremely high security, such as finance, medicine, and public services.

Background FIXER and the Okazaki Laboratory have been promoting joint research on the social implementation of LLMs since 2024. (Note) Building on this existing partnership, this initiative is positioned as a necessary step to rapidly implement rapidly evolving LLM technology into society and convert it into business value.

Reason for Adopting "GPT-OSS Swallow" While the utilization of generative AI significantly impacts corporate competitiveness, many companies, especially those handling confidential information, face high hurdles in using external AI services due to concerns about information leakage risks and data sovereignty. In response to these challenges, "GPT-OSS Swallow," developed primarily by a research team from Tokyo Tech and the National Institute of Advanced Industrial Science and Technology, is an open-source license available for commercial use. It has been reported to demonstrate high performance among existing open LLMs of comparable or smaller scale in Japanese instruction-following tasks. Its open nature allows for free customization within customers' closed environments, and its high performance directly translates to improved productivity in actual business operations. FIXER has determined that these "open" and "high-performance" characteristics are key to meeting the needs of customers aiming for AI utilization in on-premise environments.

Evolution of "Sovereign GaiXer" Driven by This Research "GPT-OSS Swallow" is characterized by its Japanese performance, which is comparable to larger models even within the relatively small parameter model group of 20B (20 billion) or less. This characteristic has high affinity with the desktop and on-premise AI utilization envisioned by "Sovereign GaiXer." Through this joint research, by optimizing and integrating "GPT-OSS Swallow" into "Sovereign GaiXer," customers will be able to utilize high-performance Japanese generative AI in a secure environment under their own control. This will enhance the value provided by "Sovereign GaiXer" from being a "secure platform" to a "secure and high-performance AI utilization platform," significantly contributing to the assurance of data sovereignty and economic security through domestic technology.

Overview of Joint Research Building on previous collaborations, this joint research will proceed with the implementation and optimization of "GPT-OSS Swallow" on "Sovereign GaiXer," anticipating commercial use. FIXER will lead the effectiveness verification based on actual business scenarios, specifically evaluating practical applicability and introduction effects in highly specialized areas, including the medical field, which has been a focus of past research, as well as finance and public sectors. Tokyo Tech will provide technical advice, expert knowledge, and language model performance evaluation. FIXER will lead the product implementation, optimization, and demonstration design for "Sovereign GaiXer" and accelerate verification towards commercialization.

Through this initiative, optimizing "GPT-OSS Swallow" on "Sovereign GaiXer" and demonstrating its practicality will enhance the value of "Sovereign GaiXer" as a more secure and accurate generative AI solution in specialized fields handling confidential information. By exploring applicability in areas with high demands for confidentiality and data sovereignty, such as medical, finance, public, and manufacturing sectors, we aim to further expand our service value.

Furthermore, in light of this joint research, we have received the following comment:

“Verification in a commercial environment is an important theme for connecting the research results of large language models to social implementation. Through this joint research, we hope to deepen our knowledge from a practical utilization perspective and further expand the possibilities of generative AI utilization.”

Tokyo Tech School of Computing, Information Science and Engineering