PFN's PLaMo 3 NICT translation hits top accuracy

Preferred Networks has positioned its PLaMo translation work as a full-scratch domestic AI effort, and this interview with three team members traces how the project began. Kentaro Imajo, who handles model development, said he proposed the translation model and built the first version himself, starting the project with a casual pitch to the president rather than a formal plan. The October 2026 release of plamo-3-translate, built on PLaMo 3 NICT, is described by the team as reaching top-level translation accuracy.

The interview also highlights how the product side shaped the model's direction. Product manager Hiroki Okui said the team initially built a Chrome extension, but strong user requests led to a Firefox add-on because Firefox's built-in translation quality was low. That feedback loop between users and development appears to have guided the expansion of language support.

Looking ahead, the team says a small model that runs locally on smartphones and PCs is coming soon. After the interview, Preferred Networks released four PLaMo 3 Translate model weights, making the technology available for local use. Whether independent evaluations confirm the claimed top-level accuracy remains to be seen.