Sakana AI said on 9 October that its Japanese-adapted model, Sakana Namazu, has been adopted by Evidence Finder, a medical literature search tool for doctors built by the Japanese medical AI company IRIS. Evidence Finder answers clinical questions by searching databases including PubMed and returning answers with citations. IRIS’s own algorithms select and rank the papers, and Namazu compares and synthesises them into the answer. A Verify feature automatically checks that the cited papers actually exist, which is a direct response to models inventing references.
Sakana’s pitch for Namazu is post-training that improves Japanese instruction following, translation and handling of Japan-specific context without losing the base model’s maths, coding and general reasoning scores.
The headline number: an evaluation model built on Namazu scored 96.4% on the 120th Japanese National Medical Licensing Examination, sat in February 2026. IRIS says that is the highest among publicly available domestic foundation models designated under the METI-backed GENIAC programme, measured across exams from the 118th onward, as of 1 September 2026.
Note how tightly drawn that claim is. It covers domestic models in one government programme, not all models available in Japan, and it is IRIS’s own comparison. Sakana’s post is also explicit that exam performance reflects knowledge and reasoning and does not by itself demonstrate clinical usefulness. The announcement is on Sakana AI’s site, in Japanese.