Gemini 3 Released: Long-Document Translation and Multimodal Understanding Take Another Step

In November 2025, Google released the Gemini 3 model family, with significant improvements in long-context reasoning, multimodal understanding, and instruction following, quickly integrated into Google Translate and Workspace. For the language industry, the most watched capability is long-document handling: a million-token context window allows an entire product manual or a full regulatory submission dossier to be translated in a single request with consistent terminology.

The industry consensus on why long context matters for translation quality is well established: traditional machine translation segments text sentence by sentence, leaving terminology and anaphora consistency to post-editing; long-context models can "read the whole document before writing," keeping terminology unified and style coherent in the first pass. Our internal benchmarks show that on technical documentation and legal contracts, first-pass quality of the new long-context models has improved by roughly one review grade compared to two years ago.

Multimodal progress matters too: models can now directly read text in PDF layouts, charts, and screenshots, greatly automating "format-preserving translation" and significantly reducing manual desktop publishing (DTP) work.

One caveat: better first-pass output does not eliminate professional review. On the contrary, as machine output "reads more human," subtle errors—misplaced figures, dropped negations, mismatched terminology—become harder to spot. Our MTPE workflow has accordingly upgraded its review checklist to target the typical error patterns of AI-generated translation.