A large language model (LLM) is trained on extensive text collections spanning languages and subject areas. Some have hundreds of billions of parameters, approaching a trillion, and can write, analyze documents, generate code, and translate. section 1.0
Broad capability can be inefficient for specialized work. For example, extracting metrics from legal documents and writing client reports require different strengths. Detailed prompts can guide behavior, but do not change a model’s underlying expertise. section 1.0 section 1.1.2
A smaller language model can be adapted to specific data, tasks, and deployment constraints by changing its architecture. This differs from retrieval-augmented generation, which supplies external information, and fine-tuning, which further trains a model on task-specific data. These approaches can also complement one another. section 1.0 section 1.1.2





