← Back to news

Pre-translation got a boost in Poedit 3.9

Poedit 3.9 brings major improvements to automatic pre-translation. If you’ve used it before, you’ll immediately notice the difference.

Lightning fast automatic pre-translation

In this release, pre-translation is dramatically faster.

No, seriously, it really is quite dramatic. Jobs that previously took minutes now finish in tens of seconds. On large files the improvement is particularly noticeable, and pre-translation often completes more than 10x faster. Just look:

Poedit 3.9 pre-translation is dramatically faster Bar chart comparing full-file automatic pre-translation times in Poedit 3.8 and 3.9. The new version is substantially faster across engines and file sizes, reducing processing time from minutes to seconds. DeepL improves from 184 to 14 seconds for a 5K-word file and from 447 to 34 seconds for a 16K-word file, about a 13x speedup. Microsoft improves from 59 to 10 seconds and from 103 to 19 seconds, about a 5x speedup. GPT support is new in Poedit 3.9 and processes the same files in 42 and 63 seconds. Poedit 3.9 pre-translation is dramatically faster Full-file pre-translation time, in seconds. Lower is better. Poedit 3.8 Poedit 3.9 Speed-up 0 s 100 s 200 s 300 s 400 s game 4,650 words DeepL 184 s 14 s 13.1 × Microsoft 59 s 10 s 5.9 × GPT (new in 3.9) 42 s wp-affil 16,180 words DeepL 442 s 34 s 13.0 × Microsoft 103 s 19 s 5.4 × GPT (new in 3.9) 63 s Samples shown: two real-world files. Google Translate omitted for brevity; its performance is similar to DeepL.

To make this happen, we completely rebuilt the pre-translation architecture from the ground up. We rethought every aspect of it: the way strings are sent to the service, how they are compressed, how translations are streamed back to minimize delays, how the server batches and processes requests, everything was optimized for speed. Caching, preprocessing, smart batching, massive parallelization, you name it. The result? A noticeably smoother and much faster translation.

We’re quite proud of it.

GPT-powered translation

Poedit 3.9 also introduces support for GPT-based machine translation, joining DeepL, Google Translate and Microsoft Translator. GPT models are quite slow, so this would have been a non-starter without the new architecture. But with the speed improvements in Poedit 3.9, and significant translation quality improvements in OpenAI’s latest models, we finally felt comfortable with it: we could now ship GPT support that is good.

Unlike traditional machine-translation engines, GPT can take more context into account. It understands comments attached to strings, can follow custom instructions, and can recognize when a translation may be ambiguous. In such cases, it may even add a note explaining the translation choice and alert you to a potential inaccuracy!

This opens up interesting possibilities. You can guide the translation style with instructions (yes, even the classic “talk like a pirate” demo works), provide additional context about the translation project, or fine-tune the translation’s tone. This is particularly useful for things like game localization, where tone and voice matter and you want to steer the translation toward a specific style.

Just a note: Because GPT models are inherently slower than specialized translation APIs, they are currently available for pre-translation only. Interactive suggestions in the sidebar need to appear almost instantly, and GPT’s latency makes it impractical there.

Try it out!

If you rely on pre-translation, Poedit 3.9 will feel noticeably faster. Download the new version here and give it a try! If you have any feedback or questions, feel free to reach out.