Back to AI Glossary
Models & PlatformsLLM

Large Language Model

A neural network trained on massive text corpora that can understand and generate human language, typically with billions of parameters.

Why it matters

LLMs are the core technology behind modern AI assistants, copilots, and automation. Understanding their capabilities and limitations is essential for building reliable AI products.

What defines "large"

The term originally distinguished models with billions of parameters (GPT-3 at 175B) from smaller predecessors. Today the boundary is blurry — models like Mistral 7B punch well above their weight class. What matters more than raw size is the quality and diversity of training data, architecture choices, and post-training alignment.

Foundation vs. fine-tuned

A foundation model (or base model) is the raw pretrained LLM. Most production use cases rely on instruction-tuned variants that have been further trained to follow directions, and RLHF-aligned versions that are safer and more helpful. Fine-tuning on domain-specific data is a separate step that specializes the model for particular tasks.

Key capabilities

  • Text generation, summarization, and translation
  • Code generation and debugging
  • Reasoning and analysis (especially with chain-of-thought prompting)
  • Tool use and function calling
  • Multimodal understanding (vision, audio) in newer models

Related Terms

Token— The basic unit of text that a language model processes — typically a word, subword, or punctuation mark, roughly equivalent to 3/4 of an English word.Fine-tuning— The process of further training a pre-trained model on a smaller, task-specific dataset to improve its performance on that particular task or domain.Context Window— The maximum amount of text (measured in tokens) that a language model can consider at once — including both the input prompt and the generated output.Hallucination— When an AI model generates information that sounds plausible but is factually incorrect, fabricated, or unsupported by its training data or provided context.

On the AI Radar

Claude Opus 4— Anthropic's frontier model family. Claude Opus 4.6 leads on complex reasoning and agentic coding, with Claude Sonnet 4.6 offering near-Opus performance at significantly lower cost.DeepSeek R1— A 685B-parameter (37B active via MoE) open-weight reasoning model from Chinese AI lab DeepSeek that demonstrated reasoning behaviour can emerge through reinforcement learning alone. The R1-0528 update significantly improved performance, approaching frontier proprietary models.Llama 4— Meta's open-weight multimodal LLM family featuring MoE architecture with Scout (109B/17B active) and Maverick (400B/17B active) models. The planned 2T-parameter Behemoth remains unreleased after a confirmed benchmark manipulation scandal.Gemini 3.1 Pro— Google's frontier multimodal model family. Gemini 3.1 Pro leads benchmarks across coding, reasoning, and multimodal tasks, with native support for text, image, audio, and video.
Back to AI Glossary