Sakana AI ships Namazu-powered JP-EN-ZH translation web app

The interesting bit of this week's launches out of Tokyo isn't another frontier model claim, it's a pragmatic product move. On July 5, 2026, Sakana AI wired a new feature called Sakana Translate into its existing Sakana Chat product, pitched as a free web app for bidirectional Japanese-English-Chinese translation. What makes it worth reading past the launch post is how it is built.

The model behind it, Namazu, isn't trained from scratch. MarkTechPost reports that Sakana is post-training existing open-weight bases, including DeepSeek-V3.1-Terminus, Llama 3.1 405B, and gpt-oss-120B, and adapting them for Japanese context. The Namazu series was announced on March 24, 2026 in Tokyo, and this translator is the first consumer surface built on top of it. The stated strengths are the places general MT usually stumbles: business honorifics, cultural concepts, place names, proper nouns, and everyday context.

The web app bundles three modes on one login. Translate accepts up to roughly 5,000 Japanese characters at a time and streams the output, with history saved automatically. Proofread cleans up drafts using diff highlighting and adjusts tone, politeness, and formality. Ask lets you follow up on a translation to interrogate a specific choice, such as why one honorific over another, or what an alternative phrasing would carry. That third mode is the one that reframes this as more than a DeepL clone, because it turns the translator into a learning surface.

The honest caveats are worth naming. Sakana reports "competitive quality" using XCOMET-XL, a 3.5-billion-parameter neural metric from Unbabel that scores translation quality from 0 to 1, on WMT 2024 General Translation task data. That is a self-reported comparison, not an independent bake-off against DeepL or Google. There is no public API yet, and enterprise essentials like SSO, audit logs, and on-premises deployment are on the roadmap rather than shipped. What the reporting does not give you is the actual score deltas against incumbents, which is where anyone evaluating this for a real workflow will end up looking first.

Still, the shape of the bet is the part worth watching. A regional lab post-training open-weight bases into a language-specific product, shipping it free to end users while holding API access for enterprise, is a template other language markets could copy without training frontier models from scratch.