PDF translation
Upload a PDF to ZeroTwo, say which language and which reader the translation is for, and read the result in the chat. You choose the model and can run the same passage through a second one. The translation comes back as text, not a typeset PDF, and anything legal, medical or binding still needs a qualified reviewer.
Le présent contrat est conclu entre les Parties en vue de définir les conditions générales de la prestation de services de traduction technique, sous réserve des dispositions de l'Article 4 ci-après.
This agreement is entered into between the Parties to set out the general terms of the technical translation services, subject to the provisions of Article 4 below.
Most disappointment with PDF translation comes from expecting the wrong output. Here is the output this page stands behind, and the one it does not.
The translation arrives as text in the chat, in the order of the source. Tell the model to keep headings as headings, lists as lists and table rows as rows, and it will usually follow the structure in text form. You read it, correct it and copy it out.
A downloadable translated PDF that keeps the original columns, fonts and footnote positions. We have not documented that path here, so check what the app offers when you try it.
Translate the text, have it reviewed, then set it into your own layout tool using the original file as the template. That is slower than a one-click export and far more predictable for forms, contracts and typeset reports.
A full translation is more than you need to decide whether a foreign document matters. Summarize the PDF instead of translating it in full, then translate only the sections that count.
We have not run a published input and output pair for this page, so there is no sample to show you. This is the check we would run ourselves, and it takes minutes.
Upload size, page-count and OCR limits are not published on this page. The test finds your limit on your file instead of trusting a number.
Choose one text-heavy page and one with a table, footnotes or a signature block. A clean page tells you nothing about the pages that will go wrong.
State the target language and who will read it: "plain English for a non-lawyer" and "formal legal English for counsel" produce different drafts of the same clause. Register matters as much as the language pair.
Have the model transcribe the source text before translating it. An OCR slip in the transcript becomes a confident mistake in the translation, and it is far easier to spot in the source language.
Send the same two pages to another model. Where the drafts agree you can skim; where they differ is where your review time belongs. Ask each model to list the terms it was unsure about, and treat that list as a prompt for your own check, not as proof.
Numbers, dates, names, defined terms and negations first. Only then paste in the rest of the file, in sections if it is long.
No language-pair test sits behind this table. It matches document types to what usually goes wrong and suggests where to start. The right model is whichever needs the fewest corrections on your own pages. See the full list on the models page.
| Document type | What usually goes wrong | Models to compare first | What to check |
|---|---|---|---|
| Contracts and legal clauses | Defined terms drift between occurrences; a shifted modal verb changes an obligation. |
| Capitalised terms stay consistent; shall, may and must keep their force; cross-references such as Article 4 still point to the right place. |
| Medical and clinical text | Units, dosages and abbreviations are expanded or converted incorrectly; negations are lost. |
| Every number with its unit, every abbreviation, every not and no. A clinician reviews the final version. |
| Scientific papers | Multi-column reading order is scrambled; symbols and equations are mangled; citations get translated. |
| Reading order of columns, symbols, figure captions, and that the reference list is left untouched. |
| Marketing and voice-sensitive copy | A literal translation keeps the words and loses the tone; idioms come across as nonsense. |
| Register, idioms, and brand or product names that should stay in the source language. |
| Chinese, Japanese and Korean pairs | Personal names, honorific level and technical terms are rendered inconsistently. |
| Names on first and later mentions, politeness level, and any term with an established translation in your field. |
| European official documents | Institutional terms have fixed official translations that a model may paraphrase. |
| Compare institutional terms against the issuing body's own multilingual glossary where one exists. |
Editorial shortlist from ZeroTwo, which publishes this page and sells access to these models. Model names are as listed in the ZeroTwo app; availability depends on your plan.
We do not publish a supported-language count, because coverage depends on the model you pick. What we can say is how much human review to plan for. These are general expectations, not measurements.
English with Spanish, French, German, Italian, Portuguese, Japanese or Chinese. Expect a readable draft. You still check numbers, names and defined terms, because fluent output hides small errors well.
Expect more literal phrasing and more inconsistency in names and terms. Run a second model, and have someone who reads the target language look at the whole draft, not a sample.
Treat the output as a lead for a reader, not as a draft. Quality varies a lot by model and by language, and no test on this page covers it. For anything that matters, use a human translator and use the model to help you follow what they changed.
Once the draft is sound, change its tone with the AI rewriter and tidy the wording with the AI grammar checker.
Free is enough to run the two-page test on one model. Comparing several models, or translating long documents, uses more of a paid plan's monthly credit pool, because heavier models and longer tasks cost more credits.
Free also has a limited model selection and no card is required. Full allowances are on the pricing page.
This page does not promise one. The translation comes back as text in the chat, and you can ask the model to keep headings, lists and table rows as text structure. If you need a file that matches the original layout, translate the text, have it reviewed, and set it into your own layout tool. Check what the app offers when you try it, because we have not documented a layout-preserving export here.
It depends on the model and the scan. Resolution, skew and contrast all affect how well the text is read. Ask the model to transcribe the source first and read the transcript before you trust the translation. For poor scans, run a dedicated OCR tool first and paste the text in.
This page does not publish upload size, page-count or OCR limits, and the old claim of no size limit has been removed. What is documented is that the Free plan allows 1 file upload per message and that paid plans raise that. Start with two pages, then translate long documents in sections.
It can be useful for understanding a document and for preparing a first draft. It should not be the document of record. A contract you will sign, a regulatory filing or a clinical record needs a qualified human reviewer, whatever tool produced the draft. The model shortlist on this page comes from reasoning about what usually goes wrong, not from a language-pair test.
Google Translate and DeepL are dedicated translation services, and both have their own document workflows. Check their pages for supported file types, output formats and limits, because those change. What ZeroTwo adds is a choice of models, the ability to ask why a term was translated a certain way or to request a different register, and one chat where you can also summarize the document or draft a reply.
There is a Free plan with 100 free credits to start, 20 bonus credits every day you log in, 1 file upload per message and a limited model selection, and no card is required. Comparing several models can need a paid plan because Free model selection is limited. Pro is $29.99 per month ($26.99 per month billed annually) with 3,500 monthly credits. Model runs draw from one monthly credit pool, and heavier models and longer documents use more of it.
To produce a translation, the text of your file is processed by the model you choose, which for most models means a third-party AI provider. Read the privacy policy and your chosen provider's terms before you upload anything confidential. For sealed legal or regulated health material, follow your organisation's data-handling rules instead of relying on this page.
This page does not publish a language count. Coverage depends on the model you pick, and ZeroTwo offers 60+ models that differ by language. The practical test is to run two pages in your language pair through two models and have a reader check them.
Privacy questions are answered in full in the privacy policy.
Ask questions of a PDF instead of translating the whole thing.
Get a digest of a long PDF when you only need the main points.
Change the tone of a translated passage for a different audience.
Polish the translated text before it goes out.
Compress translated text into a short brief.
See the models you can compare on the same passage.
ZeroTwo publishes this page and is one of the tools it describes. The Free plan needs no card.