AI library

Fundamentals

Large language models

Last reviewed 4 October 2026

Overview

A large language model (LLM) is a neural network trained to predict the next token in a sequence of text. Trained on a huge corpus and then tuned to follow instructions, that one skill turns into general abilities: answering questions, writing and editing code, summarising, translating, planning, and using tools.

Modern LLMs are decoder-only transformers with billions to trillions of parameters. Frontier models also take images, audio, or video as input, and many can "think" before answering (see reasoning models).

Key concepts

  • Token: the unit of text the model reads and writes; roughly 3–4 characters of English. See tokens and context.
  • Context window: the maximum tokens (input plus output) the model can attend to at once.
  • Parameters / weights: the learned numbers in the network. Open-weight models publish them; closed models are reachable only through an API.
  • Pre-training: next-token prediction over a large text corpus; gives broad knowledge.
  • Post-training: instruction tuning and reinforcement learning that make the model helpful, safe, and good at tasks. See training and fine-tuning.
  • Inference: running the model to generate output, one token at a time.
  • Sampling: choosing each next token from the predicted distribution (temperature, top-p). Several current models fix these settings and no longer let callers change them.
  • Hallucination: fluent but false output; reduced by grounding (see RAG), tools, and better training, but never fully eliminated.
  • Knowledge cutoff: the date after which the model has no training data.

How to use it

  • Use an API or chat app for closed models; run open-weight models locally or on a host.
  • Give clear instructions and enough context; see prompting.
  • Ground answers in documents or tool results when facts matter.
  • Measure quality on your own tasks; see evals.

Current state

  • 2026-10-04: Frontier families are Claude (Anthropic), GPT-6 (OpenAI), and Gemini (Google), with 1M-token context windows now common. Leading open-weight families include Qwen, DeepSeek, Llama, Mistral, and Gemma. See models.

Open questions

  • How far will scaling post-training and test-time compute keep improving results?
  • How to make models reliably say "I don't know" instead of guessing.

Sources