ShinrAI 1.6.1
Released on
The model of the managed API since 20 October 2026, under the model ID v1.6. The 21 data categories do not change. Against 1.3: real-world field bench 50 → 139 of 186 values, encyclopedic recall in eight languages 34.5 → 81.9, long-record redaction 36.5 → 57.6, 15-locale median 95.0 → 95.5. The 1.6 rules for names on signature lines, middle initials and Polish “r.” dates stay. Span edges are cleaner than in 1.6. A span stops at a line break. Names, addresses, contact details and identifiers lose an edge period, bracket or quote. A Japanese or Korean date stays whole. A name stops at a conjunction or a genitive, loses a Japanese honorific, keeps Portuguese particles such as “da” and “dos”, and grows to the whole word in more languages. Fragments of one or two characters inside a word are dropped. The detection threshold stays at 0.35.