Sakana AI Launches Sakana Translate, a Namazu-Powered Japanese–English–Chinese Translation Tool With Translate, Proofread, and Ask Modes
Sakana AI has added a new feature called Sakana Translate to its chat service, Sakana Chat. It handles bidirectional translation across Japanese, English, and Chinese. The translation engine is Namazu, the company’s model series adapted for Japanese.
Sakana Translate ships as a free web app. A single account unlocks all three of its modes.
What is Sakana Translate
Sakana Translate is a browser-based translation product, not a new base model. It runs on Namazu, Sakana AI’s Japan-adapted model series.
The concept Sakana AI states is ‘deep translation for Japan.’ The goal goes beyond swapping words and sentence structures. It aims to carry context, tone, and register between languages.
That focus targets a specific gap. Sakana AI argues that general tools often miss what makes Japanese distinctive. Examples include business honorifics, culturally specific concepts, abbreviations, and internet slang. Grammar may stay correct while interpersonal tone gets lost.
The product bundles three functions into one screen: Translate, Proofread, and Ask.
The Three Modes, Explained
Each mode targets a different everyday task. The table below summarizes them.
A few terms are worth unpacking for engineers new to this space.
Streaming output means the translation appears progressively, token by token. You do not wait for the full result before reading. This mirrors how chat models return text.
Diff highlighting shows exactly what changed. Additions and removals are marked inline, like a version-control diff. Proofread goes past grammar. It also tunes naturalness, politeness, and the register a reader expects.
Ask removes the tool-switching problem. You no longer jump between a translator and a dictionary. Nuance questions get answered against the same source and output.
You can try all three modes in the interactive demo below
How Namazu Powers It
Namazu is the engine, so its design matters here. Sakana AI is a Tokyo lab founded in 2023 by David Ha and Llion Jones.
Namazu is not trained from scratch. It applies post-training to existing open-weight foundation models. Reported base models include DeepSeek-V3.1-Terminus, Llama 3.1 405B, and gpt-oss-120B.
Post-training means adapting an already-trained model with further tuning. It is cheaper and faster than pre-training a model from zero. Sakana AI uses it to fit models to Japanese language and culture.
Sakana AI first announced the Namazu series on March 24, 2026. Sakana Translate applies that same adaptation work to the translation problem.
Benchmark and Performance
Sakana AI team evaluated translation quality with a standard setup. It used XCOMET-XL on the WMT 2024 General Translation task data.
Here is what those two names means:
WMT 2024 General Translation is a shared task from the Conference on Machine Translation. Systems translate test sets drawn from several domains across many language pairs. It is a common yardstick for machine translation research.
XCOMET-XL is a neural evaluation metric from Unbabel, with roughly 3.5B parameters. It is a learned model that scores translation quality. It outputs a score and also flags specific error spans. Scores run from 0 to 1, where higher is better.
By the reported results, Sakana Translate landed in a score band close behind the leading models. Sakana AI describes this as competitive quality for a translation engine.
Sakana AI also ran a qualitative check on everyday Japanese texts. It reports strengths in honorifics, cultural concepts, place names, proper nouns, and everyday context.
Use Cases With Examples
Sakana AI published two concrete outputs. Both show the tone-preservation goal in action.
1. Business email (Japanese → English). The source is a polite, indirect request. It uses set business phrases such as o-mitsumori haiken shimashita and katte na onegai desu ga.
Source: お見積り拝見しました。勝手なお願いですが、もう少しご相談できますか？ご予算あれば社内で調整しますので、お聞かせください。
Sakana Translate: I saw your quote. This is a bit of a selfish request, but could we talk a bit more? If you have a budget in mind, I can work on it internally, so please let me know.
Sakana AI notes the polite tone is preserved. A flatter translation would drop that register.
2. Internet slang (English → Japanese). The source uses casual shorthand from a group chat.
Source: Iykyk, honestly. It’s an inside joke from the group chat.
Sakana Translate: まあ、わかる人にはわかるよね。グループチャットの内輪ネタだから。
The output keeps the same casual temperature in conversational Japanese.
Beyond these, practical scenarios follow naturally. A support agent can translate a long client thread in one paste. A developer can proofread an English release note before publishing. A learner can ask why a phrase carries a certain tone.
Comparison: What Sakana Translate Targets
The table below frames Sakana AI’s stated design focus. The left column describes common behavior in general-purpose tools. The right column reflects Sakana AI’s own claims, not a head-to-head benchmark.
A Look at the Evaluation Metric
Sakana Translate has no public API yet. Sakana AI lists API access as a future, enterprise-focused plan.
You can still reproduce the evaluation method it used. XCOMET-XL is open and runnable. The snippet below scores one translation with it.
# XCOMET-XL is gated on Hugging Face: accept its license,
# then run `huggingface-cli login` before download_model().
from comet import download_model, load_from_checkpoint
model_path = download_model(“Unbabel/XCOMET-XL”) # ~3.5B params
model = load_from_checkpoint(model_path)
data = [{
“src”: “お見積り拝見しました。”,
“mt”: “I saw your quote.”,
“ref”: “I have reviewed your quote.”, # omit “ref” for reference-free QE
}]
# Returns per-segment scores, a system score, and error spans
output = model.predict(data, batch_size=8, gpus=0)
print(output.system_score) # score from 0 to 1, higher is better
Strengths and Limitations
Strengths:
Free web app with three modes behind a single account
Bidirectional coverage across Japanese, English, and Chinese
Long-input support of about 5,000 Japanese characters, with streaming
Tone and register handling for honorifics, slang, and cultural concepts
Diff-based proofreading that goes beyond grammar
Limitations:
No public API today; API access is only on the roadmap
Benchmark figures come from Sakana AI’s own evaluation
Three languages only at launch, centered on Japanese
Enterprise features like SSO, audit logs, and on-premises are planned, not shipped
Key Takeaways
Sakana Translate adds Japanese–English–Chinese translation to Sakana Chat, powered by Namazu.
Three modes ship together: Translate, Proofread, and Ask, in one free web app.
Translate handles ~5,000 Japanese characters with streaming; Proofread shows diffs; Ask explains nuance.
Sakana AI reports competitive XCOMET-XL scores on WMT 2024 data, from its own tests.
API access, file translation, glossaries, and enterprise controls are on the roadmap.
Sources
Sakana AI, “Sakana Translate: Sakana Chat Now Supports Translation” — https://sakana.ai/translate-release/
Sakana AI on X (@SakanaAILabs) — announcement post
Sakana Translate web app — https://translate.sakana.ai
Sakana AI, “Namazu (alpha)” announcement — https://sakana.ai/namazu-alpha/
Unbabel, “Introducing XCOMET” and Unbabel/XCOMET-XL model card — https://huggingface.co/Unbabel/XCOMET-XL
WMT24 General Machine Translation Shared Task findings — https://aclanthology.org/2024.wmt-1.1/
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
