04 / RESOURCES / GUIDE
LLM models explained: what they are and how to choose one
A plain-language guide to large language models. How they work, what tokens and context windows are, the difference between model sizes, and how to pick one for a business task.
Behind every AI assistant your team uses, there is a large language model, or LLM. Claude, ChatGPT, Gemini, and Copilot all run on one or more of them. You do not need to know how they are built to use them well, but a few ideas make a big difference to the results you get and the money you spend.
What an LLM actually does
An LLM is trained on a very large amount of text. From that, it learns patterns in how language works: facts, reasoning steps, writing styles, code. When you give it some text, it predicts what text should come next, one small piece at a time, and keeps going until the answer is complete.
- Your requestPlus any documents
- Split into tokensSmall pieces of text
- Model predictsThe next token, again and again
- AnswerTokens turned back into text
That simple loop explains a lot of its behavior:
- It is very good at language tasks: summarizing, drafting, classifying, extracting, translating, and explaining.
- It can be confidently wrong. It predicts what a good answer looks like. If it lacks the facts, it may produce something plausible that is not true.
- It only knows what it was trained on, plus what you give it. It does not automatically know your prices, your clients, or today's news unless that information is provided or the tool can look it up.
Tokens: how models read and how they charge
Models do not read words; they read tokens, small chunks of text. In English, a token is often about three-quarters of a word, so 1,000 tokens is roughly 750 words.
Tokens matter for two reasons:
- Cost. When a business uses a model through an API, it is usually billed per token, for both what you send and what the model writes back.
- Limits. Every model has a maximum number of tokens it can consider at once.
Context windows: the model's working memory
The context window is how much text a model can consider in one go: your instructions, the conversation so far, any documents, and its own answer. Modern models have large context windows, enough for long reports or many documents at once.
What fills a context window
- InstructionsWho it is, what to do, rules to follow
- Documents and dataFiles, records, search results
- Conversation so farEvery earlier message
- Room for the answerWhat it writes back
Bigger is not always better. Sending a model everything "just in case" costs more and can make answers less focused. Good AI systems send the right information, not all of it. That is the idea behind knowledge bases and skills (see related reading below).
Model sizes and families
Most AI companies offer a family of models in different sizes.
| Large models | Mid-size models | Small models | |
|---|---|---|---|
| Best at | Hard reasoning, complex writing, long analysis | Most everyday business work | High-volume, simple tasks |
| Speed | Low | Medium | High |
| Cost per task | Highest | Moderate | Lowest |
| Example uses | Strategy drafts, tricky code, research | Emails, summaries, reports | Tagging, sorting, extracting fields |
The main families you will meet:
- Claude (Anthropic)
- GPT (OpenAI, the models behind ChatGPT)
- Gemini (Google)
- Llama and other open-weight models, which a business can run on its own servers
Names and version numbers change often. The pattern (large, mid-size, and small models from each provider) has stayed stable.
Reasoning models and "thinking"
Some models can take time to work through a problem step by step before answering. This is often called reasoning, thinking, or extended thinking. It improves results on multi-step problems such as analysis, planning, math, and code, at the cost of speed and tokens. For a quick email draft, it is unnecessary. For "find the inconsistencies across these five contracts," it is worth it.
Where models run
| How you use it | Good for | |
|---|---|---|
| Chat apps | Claude, ChatGPT, Gemini in a browser or app | Individuals drafting, researching, analyzing |
| Inside your tools | AI features in your CRM, email, or documents | Teams who want AI where they already work |
| Through an API | Your own app or automation calls the model | Custom workflows, automation, products |
| Self-hosted | Open-weight models on your own servers | Strict data-control requirements |
How to choose a model for a business task
- Start with the task, not the model. Write down exactly what goes in and what should come out.
- Try a mid-size model first. It handles most business work well.
- Test with real examples, ten or twenty of them, and compare outputs side by side.
- Move up a size only if results are not good enough; move down if a small model does the job and volume is high.
- Check the privacy terms. Business and API plans typically offer stronger data protections than free consumer accounts. Confirm how your data is handled before sending anything sensitive.
Common mistakes
- Trusting without checking. Verify facts, figures, and anything sent to customers.
- Pasting sensitive data into personal accounts. Use your business's approved tools and plans.
- One giant prompt for everything. Break big jobs into steps; each step is easier to check.
- Choosing by headlines. Benchmarks change monthly. Your own test on your own task is what counts.
Common questions
Will an LLM learn from our data?
Not from what you send by default on most business and API plans, but terms vary. Read the provider's data-use policy for the plan you actually use.
Does it know about recent events?
Only up to its training cutoff, unless the tool can search the web or you provide current information.
What is the difference between an LLM and an "AI agent"?
An agent is an LLM given tools (search, files, apps) and allowed to take several steps toward a goal. The LLM is the reasoning engine inside it.