Large Language Model (LLM)
A large language model (LLM) is a transformer-based neural network with billions of parameters trained on internet-scale text to model and generate human language, capable of performing diverse tasks through natural-language prompts without task-specific retraining.
A large language model is a transformer-based neural network pretrained on massive text corpora — commonly hundreds of billions to several trillion tokens drawn from the web, books, code repositories, and scientific literature. The defining characteristics are parameter count (ranging from a few billion to hundreds of billions or more) and the emergent ability to perform tasks described in natural-language instructions, including tasks unseen during training. This generalization across domains, sometimes called 'in-context learning,' distinguishes LLMs from earlier task-specific NLP models.
LLMs are pretrained by predicting the next token in a sequence (autoregressive or causal language modeling) or by recovering masked tokens (as in BERT-style masked language modeling). Pretraining is followed by alignment: instruction tuning on curated prompt-response datasets teaches the model to follow instructions, and reinforcement learning from human feedback (RLHF) or related methods (RLAIF, DPO) steers outputs toward helpfulness and away from harmful content. The computational cost of pretraining frontier models runs into tens to hundreds of millions of dollars, requiring clusters of tens of thousands of GPUs or AI accelerators.
LLMs have made conversational AI, code generation, document summarization, and multilingual translation accessible through consumer products: ChatGPT (OpenAI, launched November 2022), Claude (Anthropic), and Gemini (Google DeepMind) each accumulated hundreds of millions of users within their first year. Enterprises deploy LLMs as backends for retrieval-augmented generation (RAG) pipelines, autonomous agents, customer service automation, and software development assistants such as GitHub Copilot.
By 2026, the LLM landscape is intensely competitive, with frontier models from OpenAI, Anthropic, Google, Meta (Llama family), Mistral, and several national initiatives. Context windows for leading models exceed one million tokens, enabling analysis of book-length documents in a single call. Key research directions include multimodality (integrating vision, audio, and video), improved reasoning on mathematics and formal logic, and inference efficiency — achieving near-frontier capability at substantially lower compute and energy cost.