Sakana AI Unveils Translation Tool Targeting Japanese Language Nuances

The rapid evolution of large language models has intensified focus on specialized tools that address linguistic and cultural gaps often overlooked by general-purpose systems. Sakana AI, a Tokyo-based laboratory founded in 2023 by David Ha and Llion Jones, has introduced Sakana Translate as an addition to its Sakana Chat service. This browser-based application supports bidirectional translation among Japanese, English, and Chinese while emphasizing preservation of tone, register, and context.

Sakana Translate Enhances Cross-Language Communication

Sakana Translate operates as a free web application powered by the Namazu model series, which Sakana AI adapted specifically for Japanese language and culture through post-training on existing open-weight foundation models such as DeepSeek-V3.1-Terminus, Llama 3.1 405B, and gpt-oss-120B. The product integrates three modes on a single screen and targets everyday professional and conversational tasks.

Core Features and Operational Modes

The application bundles three distinct functions designed for practical use cases.

• Translate mode converts pasted text between the supported languages, handling up to approximately 5,000 Japanese characters with streaming output and automatic history saving.
• Proofread mode refines drafts by adjusting tone, politeness, and formality, displaying changes through diff highlighting.
• Ask mode responds to follow-up questions about translation results, clarifying nuance, suggesting alternatives, or explaining grammar within the original context.

These capabilities aim to reduce the need to switch between separate translator and dictionary tools.

Technical Foundation and Performance Evaluation

Namazu applies post-training to adapt base models for Japanese-specific elements including business honorifics, culturally specific concepts, abbreviations, and internet slang. Sakana AI first announced the Namazu series on March 24, 2026. Evaluation used the XCOMET-XL metric on WMT 2024 General Translation task data, where Sakana Translate achieved scores described as competitive with leading models. Qualitative checks on everyday Japanese texts highlighted strengths in handling honorifics, place names, proper nouns, and contextual tone. Sakana AI states that general-purpose tools frequently produce grammatically correct outputs that lose interpersonal register. The new tool focuses on carrying context and tone across languages.

Strengths, Limitations, and Future Plans

Reported strengths include single-account access to all modes, bidirectional coverage of three languages, long-input support with streaming, and diff-based proofreading. Limitations noted are the absence of a public API at launch, reliance on internal benchmark results, restriction to three languages, and pending enterprise features such as single sign-on and audit logs. API access and additional capabilities like file translation and glossaries remain on the development roadmap. Sakana Translate currently functions without a public API, with enterprise-oriented access listed as a future plan. How do you see this shaping your industry? Fact Check

• Sakana AI added Sakana Translate to Sakana Chat for Japanese-English-Chinese translation using the Namazu model series.
• The free web app provides Translate, Proofread, and Ask modes in one interface.
• Namazu uses post-training on models including DeepSeek-V3.1-Terminus, Llama 3.1 405B, and gpt-oss-120B.
• Performance was assessed via XCOMET-XL on WMT 2024 General Translation data.
• Sakana AI was founded in Tokyo in 2023 by David Ha and Llion Jones, with Namazu announced March 24, 2026.