Large Language Model

A Large Language Model (LLM) is a large Neural Network built using Deep Learning to model, process, and generate human language, trained primarily on extensive Datasets of text and code represented as Tokens; Multimodal Models may also learn from images, audio, or other data. The Transformer is the dominant architecture for modern LLMs, using Self-Attention to relate tokens within a context, although other architectures are possible. Many prominent models have billions or more Parameters, but there is no universally fixed size threshold, and some models described as LLMs are substantially smaller; Mixture of Experts designs activate only a subset of their parameters for each token. Autoregressive Language Models commonly undergo Pretraining through Self-Supervised Learning with Next-Token Prediction, while other models use objectives such as Masked Language Modeling; subsequent training may include Supervised Fine-Tuning and Reinforcement Learning. Major families include GPT from OpenAI, Claude from Anthropic, Llama from Meta, and Gemini from Google and Google DeepMind, with some families including multimodal models. Applications include Text Generation, Machine Translation, Code Generation, and tasks requiring Reasoning, although successful task performance does not establish human-like reasoning, and claims of Emergent Abilities appearing discontinuously with scale remain debated. Training and running many frontier-scale models require substantial computational resources, but requirements vary with model size, task, Quantization, and deployment choices; NVIDIA is a major supplier of AI Accelerators, not a required hardware provider.