Should you build or buy Machine Translation (MT)?
Machine Translation (MT) software converts text between languages automatically, through an API that takes source text and a target language and returns a translation. Companies use it to localize product listings, support content, and user-generated text at volume, typically metered by the character and priced by providers like DeepL, Google, and Microsoft Translator.
Copy reviewed 2026-10-04 · Research revision 2026-09-25
Sources: swissmarketplace.group; arxiv.org; deepl.com; cloud.google.com; azure.microsoft.com
Can we replace DeepL or Google Translate with our own model? Yes, for the common language pairs at volume. A named marketplace runs a self-hosted model for all its traffic at a fraction of the API cost, and nothing about translation differentiates your business. Open models aren't at parity everywhere: an independent benchmark found the best locally run models still trail the top commercial engines. Validate your own pairs on representative content, then decide on volume: APIs still win on time-to-launch and rare languages.
Build it, buy it, or bridge?
When building makes sense
Building machine translation stopped being a research project once small open-weight models got close to commercial APIs on common pairs. SMG Swiss Marketplace Group documented replacing DeepL on its Ricardo marketplace with a self-hosted GemmaX2-28-2B model served through vLLM on Google Cloud, and brought monthly translation infrastructure cost down from over 7,000 Swiss francs to just under 1,000, a savings of nearly 87 percent. The team validated it with an A/B test that found a conversion difference of 0.22 percentage points, not statistically significant, before rolling the system out to all web and mobile traffic. GemmaX2 has 2 billion parameters, runs inference on a single T4 GPU, and in SMG's own testing outperformed DeepL in some cases. None of this means every workload should move off an API; it means the option is production-proven for teams with predictable volume and the capacity to run a GPU service.
When buying makes sense
Buying the API is still the default worth defending for workloads that never reach meaningful, predictable volume. Microsoft Translator charges $10 per million characters, Google Cloud Translation $20, and DeepL's Growth API $27.50 per million past the characters bundled in its plan. All three work at any volume from the first request, and cover rare language pairs and edge cases that a self-hosted open model may not have been tuned for. Ricardo's own case for self-hosting only closed after a validated A/B test at real production volume; low-volume or unpredictable workloads rarely reach that break-even point once GPU hosting, model validation, and on-call engineering are counted against the API bill. Keep the API when your volume is modest, your language pairs are uncommon, or your team has no appetite for running inference infrastructure. Translation quality itself creates no competitive edge either way, so the right call is whichever option costs less at your actual volume.
The desk read
Build-versus-buy analysis for Machine Translation (MT) is being written. In the meantime, the framework that drives every B4 call is on the B4 Index page.
Frequently asked
Has anyone actually replaced a commercial translation API with a self-hosted model in production?
Yes. SMG Swiss Marketplace Group replaced DeepL on its Ricardo marketplace with a self-hosted GemmaX2 model, cut monthly translation infrastructure costs from over 7,000 Swiss francs to under 1,000, and validated it with an A/B test before rolling it out to all web and mobile traffic.
When does buying an MT API make more sense than self-hosting?
When your volume is modest or your languages are uncommon. DeepL, Google, and Microsoft list rates between $10 and $27.50 per million characters and work immediately at any volume, which beats standing up and validating a GPU service for traffic that never reaches meaningful scale.
When does self-hosting machine translation make sense?
Once your volume is predictable and large enough that infrastructure cost beats the metered bill, and you have engineering capacity to run and monitor a GPU service. Ricardo's own team validated quality with an A/B test before committing.
How close are open-weight translation models to DeepL on quality?
Close on common pairs, with caveats. In SMG Swiss Marketplace Group's testing, the 2B-parameter GemmaX2 scored similarly to DeepL on average BLEU and beat it in some cases, and Ricardo's production A/B test showed no statistically significant conversion difference after the switch. An independent benchmark of locally run open models found the best ones still trail the top commercial engines, so test your own language pairs.