Sakana AI’s Japanese-focused large language model (LLM), Sakana Namazu, has been adopted by Iris Co. for Evidence Finder, its literature search service for physicians. The service searches databases such as PubMed in response to doctors’ questions and generates answers with citations. Its Verify feature automatically checks whether cited papers exist, helping physicians reach primary sources more quickly.

Iris’s proprietary algorithm selects relevant papers, while Sakana Namazu compares and synthesizes the literature and generates answers. The service combines Iris’s search and verification technology with Namazu’s Japanese-language capabilities in physicians’ workflow for checking reliable primary sources. Iris evaluated the model’s ability to answer in Japanese and its performance in the medical field before adoption.

Evidence Finder’s evaluation model scored 96.4% on the 120th National Medical Practitioners Examination, held in February 2026. Iris said this was the highest score among domestic foundation models designated by Japan’s Ministry of Economy, Trade and Industry under the GENIAC program, based on publicly available information. The comparison covered exam results from the 118th examination onward that could be confirmed in company materials and other public sources; the research date was September 1, 2026. Exam scores show one aspect of knowledge and reasoning, but do not directly measure usefulness in clinical practice. Iris said it also qualitatively assessed actual answers and judged the model useful for professional work.

Sakana AI says further training open models for Japanese can cause catastrophic forgetting—the loss of previously learned knowledge and reasoning—or excessive optimization for a narrow task. The company says Namazu uses its post-training technology to retain the base model’s mathematical reasoning, general knowledge and reasoning, and coding abilities on major benchmarks, while improving adaptation to Japanese language, culture, and work contexts, as well as neutral responses. It says Namazu outperformed its base model on evaluations of Japanese instruction following, translation, and Japan-specific contexts.

Sakana AI also says it is developing infrastructure that can complete inference within Japan, aiming to offer an AI platform that domestic companies and organizations can more readily choose, including in how data is handled and systems are operated. Beyond healthcare, it identified human resources, marketing, customer support, and internal research as potential uses aligned with Japanese systems and business practices. The company envisions applications that understand Japanese instructions and documents, then consistently search, organize, and explain information. It plans to broaden Namazu’s applications through balancing quality and cost, working with experts and service providers, and sharing knowledge from users’ workplaces.

Sakana AI’s Japanese-focused large language model (LLM), Sakana Namazu, has been adopted by Iris Co. for Evidence Finder, its literature search service for physicians. The service searches databases such as PubMed in response to doctors’ questions and generates answers with citations. Its Verify feature automatically checks whether cited papers exist, helping physicians reach primary sources more quickly.

Iris’s proprietary algorithm selects relevant papers, while Sakana Namazu compares and synthesizes the literature and generates answers. The service combines Iris’s search and verification technology with Namazu’s Japanese-language capabilities in physicians’ workflow for checking reliable primary sources. Iris evaluated the model’s ability to answer in Japanese and its performance in the medical field before adoption.

Evidence Finder’s evaluation model scored 96.4% on the 120th National Medical Practitioners Examination, held in February 2026. Iris said this was the highest score among domestic foundation models designated by Japan’s Ministry of Economy, Trade and Industry under the GENIAC program, based on publicly available information. The comparison covered exam results from the 118th examination onward that could be confirmed in company materials and other public sources; the research date was September 1, 2026. Exam scores show one aspect of knowledge and reasoning, but do not directly measure usefulness in clinical practice. Iris said it also qualitatively assessed actual answers and judged the model useful for professional work.

Sakana AI says further training open models for Japanese can cause catastrophic forgetting—the loss of previously learned knowledge and reasoning—or excessive optimization for a narrow task. The company says Namazu uses its post-training technology to retain the base model’s mathematical reasoning, general knowledge and reasoning, and coding abilities on major benchmarks, while improving adaptation to Japanese language, culture, and work contexts, as well as neutral responses. It says Namazu outperformed its base model on evaluations of Japanese instruction following, translation, and Japan-specific contexts.

Sakana AI also says it is developing infrastructure that can complete inference within Japan, aiming to offer an AI platform that domestic companies and organizations can more readily choose, including in how data is handled and systems are operated. Beyond healthcare, it identified human resources, marketing, customer support, and internal research as potential uses aligned with Japanese systems and business practices. The company envisions applications that understand Japanese instructions and documents, then consistently search, organize, and explain information. It plans to broaden Namazu’s applications through balancing quality and cost, working with experts and service providers, and sharing knowledge from users’ workplaces.