Machine translation

Machine translation is software that translates text, speech or text inside images from one language to another. The open models of today usually start from a general model and are then trained only to translate.

  • The basics
  • Updated

Older translation systems were built for the job from the start. The open translation models in this catalog begin with a general language model and add translation training. TranslateGemma starts from Gemma 3, Index-Translate from Qwen3.5 and PLaMo 3 Translate from PLaMo 3 NICT. TranslateGemma was fine-tuned on human translations and translations written by Gemini, then trained with reinforcement learning against reward models that score translation quality. PLaMo 3 Translate used translations synthesized with Gemma 4, followed by two rounds of preference optimization.

A few terms come up on every model card. The source and target language form the pair being translated, and a model is usually better at some pairs than others. TranslateGemma was evaluated on 55 languages, Hy-MT2 lists 33, PLaMo 3 Translate 53 and Index-Translate 150. Low-resource languages have little training text and are the hardest. Off-target output means the model answered in a language other than the one requested. Most of these models only translate, so use a general model for anything else. Texts longer than the model's input have to be split, as with TranslateGemma's 2K tokens or PLaMo's 32,768 tokens for input and output together. A model can also add or drop content, so have a person check anything important.

Related terms

  • Translation metrics Translation metrics are the scores that model cards use to rate a translation. Some count overlap with a human reference, others use a trained model or human reviewers, and their scales cannot be compared.
  • Fine-tuning Fine-tuning continues a pretrained model's training on your own examples so that it behaves better for a specific task, style or domain.
  • Reinforcement learning (RL) Reinforcement learning trains a model by rewarding good outcomes, such as a correct answer or passing tests, instead of copying example text.
  • Context window The context window is the maximum amount of text, measured in tokens, that a model can take into account at one time, including your prompt and its own answer.

Keep reading

All terms