GLOSSARY

Large Language Model (LLM)

An LLM is a transformer trained on massive text corpora — Claude, GPT, Gemini, Llama — that generates, comprehends, and uses tools in natural language.

Last updated:

Quick answer
A large language model (LLM) is a neural network with billions of parameters trained on massive text corpora to predict the next token in a sequence. Modern LLMs — GPT, Claude, Gemini, Llama — are transformer-based and power chat assistants, code copilots, retrieval systems, and agentic applications. Enterprise deployment spans API access, private cloud deployment, and self-hosted open-weight models.

WHAT IT IS

Leading LLM families include Anthropic's Claude, OpenAI's GPT series, Google's Gemini, Meta's Llama, and Mistral's open models. Models are described by parameter count, context window length, and capability benchmarks (MMLU, HumanEval, GPQA). Open-weights models can be self-hosted; proprietary models are typically API-consumed with managed guardrails.

HOW IT WORKS

Production LLM applications add structure around the raw model: system prompts, retrieval-augmented generation (RAG) for grounding, tool/function calling, evaluation harnesses (golden datasets, LLM-as-judge), content safety, and cost/latency monitoring. A good LLM product is 90% the plumbing, 10% the model choice.

WHEN TO USE

Deploy an LLM when the workload is language-heavy, pattern-rich, and tolerates probabilistic output with human review. Avoid it for deterministic rules where traditional software is cheaper and safer.

RELATED

SOURCES

Related questions.

What is a large language model?
A large language model (LLM) is a neural network with billions of parameters trained on vast corpora of text (and increasingly code, images, and audio) to predict the next token in a sequence. Modern LLMs are transformer-based and power chat assistants, code copilots, retrieval systems, and agentic applications.
What is the difference between GPT, Claude, Gemini, and Llama?
GPT is OpenAI's family. Claude is Anthropic's. Gemini is Google's. Llama is Meta's open-weight family. They differ on training data, post-training methods (RLHF, constitutional AI, etc.), context window size, tool-use capability, pricing, and availability. Most enterprises use more than one.
How large is 'large' in practice?
Current frontier models have 100B+ parameters (often much more, though many labs no longer publish counts). Open-weight models range from 1B–400B+ parameters. Size correlates with capability, but post-training technique, data quality, and retrieval now matter at least as much as raw parameter count.
What are the enterprise deployment options?
API access from the model provider (OpenAI, Anthropic, Google, Mistral, AWS Bedrock), private deployment inside a customer VPC (Anthropic via Bedrock, Gemini via Vertex), or self-hosted open-weight models (Llama, Mixtral, Qwen). Each trades cost, latency, control, and compliance differently.
How does NUUN AI choose models for client builds?
We select per use case, not per vendor loyalty. We evaluate on a held-out test set drawn from the client's actual workload, and we design architecture so the model is swappable. Vendor lock-in at the model layer is a liability we avoid by default.

Need this term in action?

Book a call