Models
Large language models
Machine translation models are machine learning models for generating a translation.
APIs provide convenient access to underlying models. Models can also be trained and deployed directly.
Evolution
With statistical machine translation, systems required multiple models:
With neural machine translation, that was replaced by one end-to-end model. Increasingly, one model is used for many language pairs.
Unlike machine translation APIs or apps, raw machine translation models typically still do not handle basic requirements of a production system, like language identification, segmentation or bridging.
Each model typically consists of multiple model versions, that vary by size or release date.
More types
In addition to neural machine translation models, there are various types of models currently used by machine translation APIs and related APIs.
- Generic generative AI (chat) models, that require prompting to generate translations
- Translation-focused generative AI (chat) models
- Quality evaluation models
- Quality estimation models
- Automatic post-editing models