Start with two candidates and one real document
Choosing a model to translate documentation is less about finding a universal winner and more about finding a dependable fit for your material. A setup guide, an API reference, and a troubleshooting article have different failure points. Your shortlist should reflect the documents you actually publish.
For a manageable first run, consider Claude Sonnet 5.5 and Gemini 3.1 Pro as two candidates to evaluate. This is a starting shortlist, not a claim that either will produce the best translation for every language. Test both on the same sample, with the same instructions, before committing to a larger batch.
Build a sample that exposes documentation mistakes
Pick a short, representative excerpt rather than your easiest paragraph. Include a procedure, a warning, a code example, and a few product terms. If your documentation contains tables or cross-references, include those too. The goal is to reveal problems before they spread across an entire documentation set.
- Procedures: Does the translation preserve the order of actions and prerequisites?
- Warnings: Does it retain conditions, negation, and the seriousness of the consequence?
- Technical strings: Are commands, parameter names, paths, and placeholders unchanged?
- Terminology: Does the same concept receive the same approved translation throughout?
Ask someone fluent in the target language and familiar with the subject to review the results. Smooth prose alone is not enough: an elegant sentence can still reverse an instruction.
Give both models the same translation contract
A translation contract is simply a clear list of what the model may change and what it must preserve. Specify the target language and locale, the intended reader, and the desired tone. Include a small glossary of approved terms, plus a list of names that must remain untouched.
For example, tell the model to translate explanatory prose while preserving code blocks, inline code, URLs, heading identifiers, and placeholders exactly. Ask it not to add examples, repair commands, or invent missing details. If the source is ambiguous, request a separate note instead of a silent guess.
Separate the translated document from review notes. Otherwise, helpful commentary can accidentally enter your published page. Keep your instructions identical for both candidates so that you are evaluating their output rather than differences in your prompts.
Choose by editing effort, not polished first impressions
Review each output against the source. Record meaning errors, glossary violations, formatting damage, and awkward language separately. A model that sounds natural but changes an authentication requirement should not win because its introduction reads well.
Also measure the work needed to make each result publishable. Count corrections and record reviewer time for your own trial. Token charges matter, but a cheaper generation can become expensive if every page needs extensive repair. Avoid assuming that your result for one language will carry over to another.
Choose a default candidate only after this review. Keep the second on your shortlist for difficult passages or future evaluation, rather than sending every document through both automatically. A second model can suggest alternatives, but agreement between models is not proof of correctness.
Run the shortlist through one relay
SwiftRelay is a practical option for trying these candidates through one pay-as-you-go relay, with per-token billing. It works with OpenAI-compatible SDKs, which can help you keep one test harness while changing the selected model. Sign up for SwiftRelay to use the small gift credit available to new accounts for an initial translation trial.
For an OpenAI-compatible SDK, use https://api.fastapirelay.com/v1 as the base URL. Keep source samples, prompts, and review criteria consistent across runs, and check current model availability and pricing before starting a larger job.
Publish only after a technical review
Once you select a default, translate a small batch and review it before expanding. Check that commands still work, references still point to the right places, and warnings still mean exactly what the source intended. Save approved terminology and corrections for the next batch. Your lasting advantage is not just a model choice; it is a translation process you can repeat and audit.