Sakana AI has launched Sakana Translate, a free web application for translating between Japanese, English, and Chinese, built on Namazu, the company's series of models adapted for the Japanese language. The tool lives inside Sakana Chat and offers three modes. A translation mode handles inputs of up to roughly 5,000 characters and streams the output in real time. A proofreading mode rewrites text into more natural phrasing and shows exactly what it changed through diff style highlighting. And an ask mode lets a user pose questions about a passage rather than just convert it. The launch was announced on July 6, 2026, and the service is available for free on the web.
The pitch is what Sakana calls deep translation for Japan. Instead of treating translation as swapping words and rearranging grammar, the tool is designed to carry the things that usually get lost, the tone of a message, its level of politeness, its register, and the sense of distance between the people involved. This matters far more in Japanese than most languages, because formality is not an optional layer of style, it is grammatically baked in, and a translation that gets the words right but the register wrong can read as rude, cold, or oddly stiff. The proofreading mode leans into the same idea, adjusting not just grammar but naturalness, politeness, and tone.
On measured quality, Sakana is careful not to oversell. The company reports that Sakana Translate scores in a band close behind the leading models on the XCOMET-XL metric, evaluated on the WMT 2024 general translation task data. In plain terms, that is competitive rather than record breaking, and the honest reading is that the product is not trying to win a benchmark race outright. Its argument is that for the kind of everyday, culturally loaded translation people actually do between Japanese and other languages, a model tuned for the job plus a free, well designed interface is worth more than a fraction of a point on a leaderboard.
The company behind it gives the launch more weight than a typical translation app would carry. Sakana AI is a Tokyo based lab co founded by Llion Jones, one of the co authors of the 2017 Transformer paper that underpins modern large language models, and David Ha, and it has spent its short life arguing for building frontier AI in Japan rather than importing it wholesale. Namazu is the concrete expression of that thesis, a model family shaped around Japanese language and context. Sakana Translate is Namazu pointed at one of the most practical, and most culturally sensitive, tasks a language model can do.
The reason a translation tool is worth noticing here is that it sits on top of a bigger question the industry keeps circling in 2026, which is whether it is better to reach for one giant general model for everything or to build smaller models tuned to a specific language, culture, or market. Translation is close to the perfect test case for that argument, because it is exactly where a one size fits all system tends to flatten the nuances that a locally grounded one can preserve. Sakana Translate will not replace the big general engines, and it is not pretending to. But as a free, focused, and openly benchmarked product from a credible lab, it is a clear and usable piece of evidence for the sovereign, culturally tuned model approach, at a moment when a lot of countries and labs are making the same bet in theory.