Design machine translation workflows with confidence.
How it works
Machine translation (MT) is still a viable option, and it comes in many forms. Some engines can be trained on parallel data, some on glossaries, and some are enhanced with AI. Where do neural machine translation (NMT) engines stand next to large language models (LLMs)?
Does NMT output quality stack up against LLMs and, if so, is it more cost-effective and energy-efficient? The risk is adopting an LLM because it's new, when a time-tested MT engine may meet your needs.
NMT engineLLM
ConsistencySame input gives the same output, run after run.Output can vary between runs unless tightly constrained.
CustomizationTrained on parallel data and glossaries.Steered with prompts, examples, glossaries, or fine-tuning.
Tone and contextSentence by sentence, closer to literal.Stronger when tone and surrounding context matter.
CostGenerally low and predictable per word.Varies with the model and the length of each prompt.
Energy useSmaller models, less compute per word.Larger models, more compute per word.
Typical failureAwkward or overly literal phrasing.Fluent text that adds or drops meaning.
General tendencies, not verdicts. Which one wins depends on your content, and that is what testing is for.
Not sure whether you need an LLM or an MT engine? Let's compare them on your content.