What Is a Large Language Model? A Plain-Language Guide for Web and Nonprofit Teams

A large language model (LLM) is a program trained on a huge amount of text to do one job: predict the next token, a small chunk of text, given everything before it. Repeat that prediction a few thousand times and you get an email draft, a summary, a block of PHP or an answer to a member’s question. ChatGPT, Claude and Gemini all run on large language models.

This guide is for agency staff, nonprofit communicators and CMS developers who use these tools every week but never got a plain explanation of what sits underneath. It covers how a large language model is trained, what tokens, context windows and parameters are, why answers change from run to run, and why models make things up. It also covers open-weight and closed models, how LLMs get built into products, their limits, and how to pick one. Model names and prices change fast, so they sit in one dated section near the end.

The short version: an LLM is a very good text predictor that was then trained to follow instructions. It does not look facts up unless you connect it to a source. It can be fluent and wrong in the same sentence. Pick one by testing it on your own work, at a cost you can sustain, under data terms your client accepts.

What a large language model actually does

Google’s Machine Learning Crash Course says a language model "estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." A large language model is the same idea at a much bigger scale.

Here is the loop. You send text. The model splits it into tokens. It scores every token it knows by how likely it is to come next. One token is picked and added to the text. Then the loop runs again, one token at a time, until the model stops.

That loop explains a lot of what you see in practice. Answers stream onto the screen piece by piece because they are written that way. Longer answers take longer and cost more. And there is no separate step where the model checks a fact. It writes what is likely to come next, given what it learned and what you gave it.

Most modern LLMs are built on the Transformer, a design introduced in the 2017 paper Attention Is All You Need by Vaswani and colleagues. The paper dropped older step-by-step designs and relied on a mechanism called attention. Attention lets each token weigh how much every other token matters to its meaning. Google’s course uses the word "it" in a sentence about an animal and a street. Attention is how the model links "it" to the right noun.

For the longer history, our Natural Language Processing 101 and Machine Learning Demystified posts cover the field LLMs grew out of.

How a large language model is trained

Training happens in stages, and each stage changes what the model is good at.

Pretraining. The model reads a huge body of text and practices predicting the next token, adjusting itself each time it gets one wrong. Anthropic’s glossary describes this as training "on a large unlabeled corpus of text." The result knows a lot about language and the world. It is also awkward to use. Anthropic notes that pretrained models are "not inherently good at answering questions or following instructions."

The 2020 GPT-3 paper, Language Models are Few-Shot Learners by Brown and colleagues, showed how far pretraining alone could go. GPT-3 had 175 billion parameters. The authors gave it tasks only as text prompts with a few examples, and no extra training. It did well on translation, question answering and word puzzles.

Instruction tuning. Next, the model trains on examples of good answers to instructions, written by people. This teaches it what a helpful reply looks like.

Reinforcement learning from human feedback (RLHF). Then people rank several answers to the same prompt, and those rankings push the model toward the answers people prefer. The 2022 InstructGPT paper, Training language models to follow instructions with human feedback by Ouyang and colleagues, set out this recipe.

What InstructGPT showed about size and tuning

The InstructGPT paper’s best-known result is about size. People preferred answers from a 1.3 billion parameter InstructGPT model over answers from the 175 billion parameter GPT-3. That is a model with 100 times fewer parameters winning on the tasks people cared about.

The authors also reported more truthful and less toxic output, with small losses on standard benchmarks. They were open about the limits. InstructGPT still made simple mistakes.

That result matters for anyone buying AI tools. How a model is trained after pretraining shapes how useful it feels, often more than raw size. Vendors now add more stages, such as training for step-by-step reasoning and tool use. The basic order still holds: read a lot of text, learn to follow instructions, then learn what people prefer.

It also explains why two models of similar size can feel very different. Their pretraining data, tuning examples and human feedback all differ. Benchmarks capture part of that. Your own test set captures the part that matters to your client.

Tokens and context windows

Models do not read words. They read tokens. A token can be a whole word, part of a word, a single character or a punctuation mark. Google’s course says one token is about three quarters of a word in English. Anthropic says one token on its recent Claude models is about 3.5 English characters. Each vendor splits text its own way, so the same paragraph gives different counts on different models.

Tokens matter for three reasons:

• Cost. API prices are per million tokens, and output tokens usually cost several times more than input. Our post on input vs output tokens breaks that down.

• Speed. Every output token is one more pass through the loop, so long answers take longer.

• Limits. Every model has a context window.

The context window is how much text a model can work with at once. It holds your instructions, any documents you paste, the conversation so far and the reply. Anthropic’s glossary describes it as the text a model "can look back on and reference when generating new text." As of October 2026, the main models from Anthropic, OpenAI and Google each offer about one million tokens.

A big window has costs. Long prompts are billed on every call, and extra text can bury the part that matters. We looked at both sides in When More Tokens Make AI Output Worse. For the basics, start with What Are Tokens?

What "parameters" means in a large language model

Parameters are the numbers inside a model that change during training. Each time the model guesses the wrong next token in pretraining, training nudges a huge number of them so the right token becomes a little more likely. After enough text, the parameters hold the patterns the model learned: grammar, facts, writing styles and code syntax.

Parameter count is a rough measure of size. GPT-3 had 175 billion. Google’s course says Transformers can contain hundreds of billions or even trillions of parameters, and that research shows larger ones consistently beat smaller ones.

Three cautions when you compare models:

• Closed vendors often do not publish a count. The model pages we checked for Claude, GPT-6 and Gemini list prices and context windows. None lists a parameter count.

• Bigger is not automatically better for your job. InstructGPT showed a small, well-tuned model can win. Small models are also faster and cheaper.

• Some models use only part of their parameters per token. Open-weight releases often give a total count and a smaller active count. Our Muse Spark Open post shows what that means in practice.

Parameters also decide hardware. The weights have to fit in memory to run. That is why open models come in several sizes, and why running one yourself starts with a memory question. Our post on hardware for AI agents covers that side.

Why the same prompt gives different answers

Ask the same question twice and you may get two different answers. That is expected.

At each step, the model produces a ranked list of possible next tokens. It does not always take the top one. A setting called temperature controls how much chance goes into the pick. Anthropic’s glossary says higher values give more varied output, and lower values give more predictable output.

Low temperature does not guarantee the same answer. Anthropic’s glossary notes that identical inputs can produce different outputs even at temperature 0. Small wording changes, the order of documents in a prompt, earlier turns in the conversation and vendor model updates all shift results too.

Some newer models no longer accept custom temperature at all. Claude Sonnet 5.5 rejects non-default values, as we covered in our Claude Sonnet 5.5 post. Anthropic’s API version 2026-10-21 goes further and rejects temperature on every Claude model, as covered in our post on the Claude API and SDK major versions.

What to do about it:

• For work that must be consistent, such as tagging posts or pulling fields from form submissions, use fixed instructions, structured output and a test set.

• For drafting, variation helps. Ask for three options and pick one.

• For anything that ships, test with many runs. One good answer does not mean every answer will be good.

Hallucination: why large language models state wrong things

A hallucination is a confident answer that is wrong. Anthropic’s guide to reducing hallucinations describes it as text that is "factually incorrect or inconsistent with the given context."

It follows from how the model works. The model writes likely text. A made-up citation in the right format, a plausible date or a policy that sounds like your client’s can all be likely text. A 2025 paper, Why Language Models Hallucinate by Kalai and colleagues, argues that training and testing reward guessing over saying "I don’t know." In the authors’ words, "language models are optimized to be good test-takers."

Anthropic’s guide lists steps that work with any vendor’s model:

• Tell the model it may say it does not know.
• For long documents, have it pull exact quotes first, then answer only from those quotes.
• Ask for a source for each claim, and drop claims it cannot support.
• Tell it to use only the documents you provide.

The guide also says these steps reduce hallucinations without removing them. For anything public, legal, financial or medical, a person checks the facts. For nonprofits, that includes grant language, donor letters and member help pages.

Open-weight and closed models

A closed model runs only on the vendor’s servers. You reach it through the vendor’s app or API. Claude, GPT-6 and Gemini are closed models.

An open-weight model publishes its parameters, called weights, for download. You can run it on your own hardware or with a host you pick. Meta’s Llama and Muse Spark Open and Google’s Gemma are examples. Google’s Gemma documentation calls the family "lightweight, state-of-the-art open models" that share research with Gemini.

Open-weight does not always mean open source. Weights can be free to download under a license with conditions, such as notices or use limits. Read the license before a client project depends on one. Our open vs closed AI models explainer sorts out the terms, and What Is Llama? covers the best-known family.

To try one, Local AI Models and What Is Ollama? walk through the setup. Running your own model also means owning its security, which our local AI agent security post covers.

How large language models show up in products

Most people meet an LLM through one of four patterns.

Chat apps. ChatGPT, Claude and Gemini put a model behind a chat window, with file uploads, search and memory built around it. The app is a product built on the model. Staff on different plans may be using different models without knowing it.

APIs. Developers send text to a model over an API and get text back. This is how a CMS plugin writes alt text, a support widget answers questions or a script tags 5,000 old posts. You pay per token. Our Vercel AI SDK walkthrough shows a basic build.

RAG. The model does not know your client’s content. Retrieval-augmented generation searches your own documents first and puts the relevant passages into the prompt. The model then answers from them. This is the usual design for "ask our knowledge base" tools. See What Is RAG? and What Are Vector Databases? for the search side.

Agents. An agent is an LLM that calls tools in a loop. It searches, reads a file, updates a record, checks the result and decides what to do next. The Model Context Protocol (MCP) is a common way to connect those tools, and our AI agents guide covers the basics.

Agents raise the stakes. A wrong prediction in a chat produces bad text. A wrong prediction in an agent can change data in your CMS or CRM.

What large language models do well, and where they fall short

LLMs are good at work built from language. That includes drafting and rewriting, summarizing long documents, and pulling names, dates and amounts out of messy text. It also includes tagging and sorting content, translation, explaining or writing code, and answering questions from documents you supply.

The limits are just as consistent:

• Knowledge cutoff. A model knows its training data up to a date. OpenAI lists April 30, 2026 for GPT-6 Astra. Anthropic lists June 2026 for its current Claude models. Anything newer needs search or RAG.

• No built-in fact check. See the hallucination section above.

• Exact math and counting. Prediction is a poor way to add a column of numbers. Have the model write and run code, or check the figures yourself.

• Instructions hidden in content. A model can treat text it reads as instructions. A web page or pasted email can steer it, which is called prompt injection. Our post on Claude Opus 5.5 prompt injection shows how much results vary by source.

• Privacy. What staff paste goes to the vendor under that plan’s terms. Check whether a plan uses data for training before anyone pastes member or donor records.

• Cost creep. Long prompts, long histories and reasoning modes all add tokens. Reasoning tokens are usually billed as output.

Large language model families, as of October 2026

This is the one section of this guide that will date. Names, prices and limits below come from each vendor’s model pages, checked on October 21, 2026. Prices are list API prices per million tokens. Chat apps are priced per seat instead.

Each vendor names a starting point. Anthropic’s models overview says to start with Claude Opus 5.5 for most workloads. OpenAI’s models page suggests GPT-6 Astra, with Sol to balance cost and Luna for high volume. Google’s Gemini models page lists Gemini 4 Argon as stable, and its pricing page shows the dated price rises above.

Two fine-print notes. Haiku 5.5 moves to $0.50 / $2.50 for prompts over 100,000 input tokens. GPT-6 Luna bills prompts over 272,000 input tokens at a higher rate. For launch detail, see our posts on Claude Haiku 5.5, Gemini 4 Argon and GPT-6 Sol and Luna.

How to choose a large language model for a project

The right model depends on the job, the volume and the data. A practical order:

1. Start from the task. Write down what the model does, how often and who sees the output. Alt text for 4,000 images is a different job from a member-facing help bot.

2. Build a small test set. Collect 20 to 50 real examples from the client’s own content, with what a good answer looks like.

3. Try three tiers. Run a top model, a mid-tier model and a small one. Small models often handle tagging and routing fine. Our Opus vs Sonnet vs Haiku guide shows how to split work by tier.

4. Price it on real traffic. Count input and output tokens per task, then multiply by monthly volume. Grade on cost per correct answer.

5. Check the data terms. Know where prompts go, whether they train future models and how long they are kept. Match that to the client’s policy.

6. Pick closed or open-weight. Closed is faster to launch. Open-weight fits when data must stay in-house and someone can run the server.

7. Plan for change. Vendors retire models on fixed dates. Keep the model ID in one setting, and re-run your test set when you switch. Our post on OpenAI’s October 2026 model shutoff shows what a retirement date looks like.

Most projects succeed or fail on the work around the model: the prompts, the data and the review step. Our post on AI implementation makes that case in more detail.

Frequently Asked Questions

What does LLM stand for?

Large language model. It is a model trained on a very large amount of text to predict the next token, then tuned to follow instructions. “Large” refers to the amount of training data and the number of parameters.

Is ChatGPT a large language model?

ChatGPT is an app built on OpenAI’s GPT models, which are large language models. The same goes for Claude and Gemini: each app wraps a model with features like file uploads, search and memory.

Does a large language model look things up on the internet?

Only when it is connected to a search tool or your documents. On its own, it answers from what it learned in training, which stops at its knowledge cutoff date.

What is the difference between a token and a word?

A token is the unit a model reads and writes. It can be a word, part of a word or a single character. Google’s course puts one token at about three quarters of an English word.

What is a context window?

It is how much text a model can work with at once, including your instructions, documents, conversation history and the reply. As of October 2026, the main models from Anthropic, OpenAI and Google offer about one million tokens.

Why does an LLM give a different answer each time?

The model picks each next token from a ranked list, with some chance involved. Temperature controls how much. Anthropic notes that output can still vary even at temperature 0.

Why do large language models make things up?

They write likely text, and a false statement can be likely text. Research by Kalai and colleagues argues that training and testing reward guessing. Grounding answers in your own documents and asking for sources reduces the problem.

Can I run a large language model on my own computer?

Yes, with an open-weight model and a tool such as Ollama. Small models run on a laptop. Large ones need a lot of memory or a rented server.

Is it safe to paste member or donor data into an LLM?

It depends on the plan’s data terms. Check where prompts are stored, whether they are used for training and what your organization’s policy allows before anyone pastes personal records.

Which large language model should my organization use?

There is no single answer. Test two or three models on 20 to 50 real examples of your own work, compare cost per correct answer and check the data terms. The dated table above lists current options.