What is an LLM: large language models, simply explained.

What is an LLM?.

An LLM (large language model) is an artificial intelligence program trained on very large amounts of text to understand and generate language. It works by predicting, step after step, the most likely continuation of a text.

It is the technology running ChatGPT, Claude, Gemini, Mistral's Le Chat or DeepSeek. When people talk about "generative AI" for text, they almost always mean an LLM.

The three letters read as follows:

• Large: the model is trained on huge volumes of text, and has a very large number of parameters, the internal settings it adjusts during training.
• Language: it handles language, human or computer (code is a language like any other).
• Model: it is a statistical model, not a database. It does not "look up" an answer, it builds one.

How an LLM works.

Without the math, an LLM always follows the same mechanics:

Before getting there, a model goes through two main phases: pre-training on huge volumes of text, where it learns language and a large amount of knowledge, then fine-tuning, where it learns to follow instructions, answer helpfully and refuse certain requests.

The context window is how much text the model can read at once. According to Anthropic, 1 million tokens is roughly 555,000 English words: enough to analyze dozens of contracts at once.

The main LLMs in 2026.

Here are the major model families and their current versions, according to official pages as of September 27, 2026:

The "published weights" column was checked on Hugging Face on September 27, 2026.

Each family has its own guide: the Claude API, Mistral Vibe, DeepSeek, and the overall comparison in Which AI to Choose in 2026.

What an LLM cannot do on its own.

An LLM is powerful, but it has limits worth knowing before you hand it a process:

• It can hallucinate. It produces the most plausible answer, not necessarily the most accurate. A figure, a reference or a quote can be confidently made up.
• Its knowledge has a date. Every model has a knowledge cutoff: June 2026 for Claude Opus 5.5, February 2025 for Claude Haiku 4.5, according to Anthropic. Beyond that, it knows nothing unless you provide the information.
• It does not know your business. Your customers, prices and procedures are not in its training data.
• It does not act. On its own, it produces text. It does not update a CRM or send an email.

Each limit has its answer: connect the model to your documents (this is called RAG), give it tools to act, and have important outputs checked by a rule or a person.

How a business uses an LLM.

There are three levels of use, from simplest to most valuable:

1. The individual assistant. Your teams use ChatGPT, Claude or Gemini to draft, summarize, translate. Immediate gains, but they depend on each person.
2. The LLM connected to your documents. An assistant that answers from your procedures, contracts or product sheets, with sources. Answers become reliable and specific to your business.
3. The LLM at the core of an agent. The model reasons, and the system around it acts: it qualifies a lead, enters an invoice, follows up with a customer. That is where time savings become significant. See What Is an AI Agent.

For the third level, the model is chosen task by task: a light, cheap model for volume, a powerful one for complex cases, a model hosted in the right region when data requires it. That is what we build in our custom AI agents.

LLM: what people ask us.