Translating a PDF usually means uploading it to an online service. For a menu or a brochure, that is fine. For a contract, a medical letter or an internal report, it means handing a confidential document to a third party. Open-weights translation models make a different approach increasingly practical: translating on your own computer.
What changed in 2026
Several capable translation models are now available with open weights and permissive licences. At the end of September 2026, Bilibili's Index team released the Index-Translate family under Apache 2.0, published on its Hugging Face page. As reported in early October 2026, the family includes text models at roughly 2B and 9B parameters and a larger mixture-of-experts preview, covering about 150 languages, plus a variant aimed at full-document translation.
Models like these can run on a reasonably modern laptop, at least in their smaller sizes.
How local PDF translation works
A PDF is not plain text. Translating one locally usually involves:
- Extracting the text. For digital PDFs, text can be read directly. For scans, OCR comes first. See PDF OCR explained.
- Translating with a local model, ideally paragraph by paragraph so context is preserved.
- Rebuilding the document, either as a new plain document or by placing translated text back into the layout.
Step 3 is the hardest. Translated text is often longer or shorter than the original, which breaks careful layouts. For most work documents, a clean translated text file or a simple re-formatted document is more useful than an exact visual copy.
Docento.app can extract text with its PDF to Text tool in the browser, without uploading, which is a convenient first step before a local model.
What hardware you need
As a rough guide, depending on quantisation:
- 2B-class models run on most laptops with 8 to 16 GB of RAM.
- 9B-class models are more comfortable with 16 GB or more, and faster with a capable GPU.
- Larger models need more memory even if only part of the model is active per word.
Expect local translation to be slower than a cloud service, but private.
Where it falls short
- Quality varies by language pair. Common pairs are strong; rarer languages vary.
- Tables and forms lose structure when text is extracted.
- Legal and medical documents still need a qualified human translator for anything official. Local AI is good for understanding a document, not for certifying a translation.
- Previews change. Check the licence and model card before relying on a preview release.
For a broader comparison of approaches, see AI PDF translation explained and how to translate PDF documents.
Takeaway
Open-weights translation models now make private, local translation of PDFs realistic on ordinary hardware. Extract the text, translate locally, and rebuild a simple document. Use it to understand and draft, and keep professional translators for anything official.
