Soft launchDialecto is still in testing, and details on this site may change as we finish it.

Translation aids

Dialecto’s aids share one stance: they propose, you decide. No aid writes a translation, stages a change, or opens a pull request on its own — everything an aid produces goes through the same staging, review, and pull-request flow as a hand-typed translation.

Machine-translation suggestions

In the editor, a suggestion is drafted for the entry you’re on and grounded in your curated context — domain, usage, brand voice, glossary, translation memory. You review it, edit it or take it, and save through the normal staging path.

Before any model is asked, Dialecto looks in your translation memory. If your organization already has an exact match for the same source text in the same language, that translation is used, and no model is called.

On your own machine

Machine translation can run on your team’s own machine. The default adapter is a local Ollama server beside the app, so your strings don’t leave the machine the app runs on. That suits a self-hosted install; Dialecto’s hosted app does not run a local model today, so a hosted team chooses DeepL or Gemini below. Which model your Ollama server hosts is your choice. Dialecto is model-agnostic and doesn’t claim a “best translation”; it makes sure whatever model you run sees the right context.

We recommend translategemma:12b through Ollama. Google’s published results show it making fewer errors than the larger general model gemma3:27b, which is still the default setting today, and it needs less memory. Those results are not measurements on your strings, and we have not yet run our own comparison, so try it on a sample of your own text. Self-hosted installs point the adapter at their own Ollama instance (see Self-hosting).

Hosted translation: DeepL or Gemini

A team without a local model can still get suggestions. During the soft launch, an owner or admin can choose hosted translation for the organization, on the organization page. It is off until you choose, and you pick one provider:

  • DeepL keeps your strings in the EU. The source text goes to DeepL SE in Cologne, Germany. DeepL’s paid service does not store it or use it for training. We use only DeepL’s paid plans and refuse its free tier for exactly that reason.
  • Gemini sends your strings to Google, a US company. The source text, and your brand voice and personas, go to Google through its paid Gemini API, which does not use them for training. Gemini follows the brand voice and personas more closely than DeepL does. If you need everything to stay in the EU, choose DeepL.

Dialecto pays the provider. Each organization gets a one-time free pass of 250,000 characters, counted by the length of each source text sent to a provider. The organization page shows how much is used and how much is left. It does not renew: when it is used up, hosted suggestions pause with a message in the editor, and you can write to info@dialecto.eu to continue. A provider we have not switched on yet appears as “Available on request”.

Placeholders such as %{name} or {count} are protected for both, and a suggestion that loses or changes one is discarded instead of being offered. Hosted suggestions work like any other: they are proposals, and they go through the same review and pull request. An exact translation-memory match is used before either provider and never counts against the free pass.

Autofill: the translation_mode setting

Per repo, you choose how far machine translation is allowed to go:

  • suggest (the default) — aids only ever propose in the editor; nothing is filled in for you.
  • autofill_submit — untranslated strings are filled with machine-translation drafts as staged changes submitted for review; a human reviews and approves each one before anything ships.
  • autofill_approve — drafts are staged and auto-approved, ready for batching into a pull request. The PR itself — and your GitHub review of it — is still the gate.

A companion switch, auto-translate on scan (off by default), runs your chosen autofill mode automatically when a scan lands new untranslated strings.

Translation memory

Dialecto remembers what your team has translated. When an identical or matching source string comes up again, the past translation is offered, so the same button label isn’t re-translated in every domain.

An exact match, the same source text for the same language in your organization, is used before any model is asked, so it costs nothing and keeps your wording. A close but not identical match is only shown as a hint beside the entry. It is never filled in for you.

Glossary

Your organization’s glossary defines the terms that must stay consistent — “Workspace” must never become “Arbeitsbereich”. Glossary terms are:

  • suggested while translating, next to the entry;
  • checked during validation — a translation that contradicts a glossary term is flagged before it ships;
  • managed at the organization level, so every repo shares one vocabulary.

Where the boundaries are

The aids never bypass review: suggest proposes only, both autofill modes stage through the review/PR flow, and the quality check is advisory unless you opt it into blocking. If you turn everything off, Dialecto is still a fully manual, fully validated editor.