Large language model
A large language model (LLM) is an AI model trained on very large amounts of text to predict and generate language. The same class of model behind general AI assistants can also translate, and recent translation systems are built on them. LLMs are highly fluent across many tasks, but they generate by prediction, so they can be confidently wrong.
How it works
An LLM is trained to predict the next piece of text given what came before, over enormous and varied data. That single objective, done at scale, produces a model that can translate, summarise, answer questions and more, often without being trained specifically for each task.
For translation, this brings strong fluency and context handling, but also the LLM's characteristic risks: it can invent detail, drift from the source, or state errors with complete confidence.
How SourceTarget uses it
Where an LLM-based engine is the best option for a given content type and language direction, SourceTarget uses it as the drafting engine, inside the same human-in-the-loop workflow as any other machine output. Its fluency does not change the rule that a human checks anything published.
Large language model compared with Neural machine translation
| Large language model | Neural machine translation | |
|---|---|---|
| Trained for | Language in general, many tasks | Translation specifically |
| Strength | Broad fluency and context | Focused, efficient translation |
| Risk | Can invent or drift from the source | Fluent but can mistranslate |