AI Translator

The 15 target languages and what to know about each

How the translation is produced

The page sends your text and the target language to the site's API, which hands it to a language model via Groq's service with a fixed instruction: translate accurately and naturally, preserve meaning, tone, names and formatting, use idiomatic phrasing, and return only the translation with no notes, no transliteration and no copy of the original. Your text is processed on that third-party service; this site doesn't store it.

Randomness is set low (temperature 0.3), so repeating a translation gives nearly the same wording each time. Because the model infers the source language from the text, you can translate from any language it understands, including into pairs that never pass through English, such as Arabic to Turkish or Russian to Spanish. Mixed-language input is handled sentence by sentence.

Size limits and why long translations can stop early

The server truncates the input at 5,000 characters, which the counter under the box flags in red, and the reply is capped at 1,500 tokens. For English-to-English-like pairs that's a comfortable margin: 5,000 characters is about 850 words, which needs roughly 1,100 tokens.

The margin shrinks for scripts that cost more tokens per word. Arabic, Hindi, Japanese and Chinese output often needs 1.5 to 3 times the tokens of the same text in English, so a 5,000-character paste translated into Hindi can hit the cap and end mid-paragraph. Keeping each request under about 2,500 characters avoids it, and a burst of more than 12 requests in a minute from one address is refused until the minute passes.

What it handles well and where it guesses

• Idioms are translated by sense rather than word by word, so “it's raining cats and dogs” becomes the local expression for heavy rain, not a sentence about animals.
• Names, brands and product codes are kept as written; the prompt forbids transliteration.
• Line breaks, bullet lists and numbered steps are preserved, so a formatted email comes back formatted.
• Formality is a guess. There's no tu/vous or du/Sie switch, so address a specific reader in the source text (“Dear Mr Dupont”) and the model usually picks the matching register.
• Gender is a guess too. Arabic and Spanish mark it on verbs and adjectives; if the reader's gender matters, say it in the source.
• Official, legal or medical documents still need a qualified human translator. Use this for understanding, drafts, messages and study.

Getting a better result from the same text

Complete sentences translate better than fragments, and abbreviations unique to your workplace will be left as-is or guessed. Expanding “ETA EOD” to “expected by the end of the day” before translating costs five seconds and removes the guess. If a term must stay in the original language, put it in quotation marks; the model generally leaves quoted terms alone.

For a back-check, translate the result into the source language in a second run and compare. Differences point at ambiguous sentences worth rewording. The Diff Checker on this site shows the two versions side by side.