A model trained to understand and generate text by learning the statistical patterns of language — the foundational class of model underpinning search, AI assistants, and every DEJAN classifier.
A language model is a model trained on large amounts of text to learn the statistical structure of language — which words follow which, how meaning is carried by context, how sentences are structured. At its core it assigns a probability to sequences of text. Modern language models use transformer architectures and are trained on web-scale corpora.
The output depends on how the model is used: a generative language model completes or continues text; a discriminative language model produces representations or scores used for classification, retrieval, or ranking. Both are language models; the distinction is in the training objective and the output head.
Language models span a wide size range. Small specialist models like DEJAN's query classifiers (ALBERT, DeBERTa) have tens or hundreds of millions of parameters and run on a single GPU. Large Language Models have billions to trillions of parameters and require distributed computing infrastructure. Distilled models sit between the two — smaller than their teachers but inheriting much of their capability.
Search engines, AI answer systems, and brand recommendation engines are all language models — or are driven by them. Understanding how language models assign probability, store factual associations, and retrieve knowledge is the foundation of AI SEO. Primary bias, associative embeddedness, and grounding snippets are all consequences of how language models work.
