What is a large language model (LLM)?
A large language model (LLM) is a neural network trained on a very large body of text to predict the next token in a sequence. A token is a word or part of a word. By learning to predict text well, the model absorbs grammar, facts, reasoning patterns, and the structure of documents such as contracts, emails, and code. The same model can then draft a letter, summarize a report, extract fields from an invoice, or answer a question, because each of those tasks can be framed as continuing a piece of text.
Almost every current LLM uses the transformer architecture, which reads all the tokens in a prompt at once through an attention mechanism. Size is measured in parameters, the learned numeric weights inside the network. Production models range from a few billion parameters, which run on one workstation GPU, to several hundred billion, which need several GPUs working together. A larger model generally knows more and reasons better, but it costs more to run and responds more slowly.
How a large language model works
Training happens in stages. Pre-training feeds the model trillions of tokens of text and adjusts its weights until its predictions improve. The result is a base model that completes text but does not follow instructions well. Instruction tuning then teaches it to respond to requests in a helpful format, and preference training such as RLHF (reinforcement learning from human feedback) shapes tone, safety, and refusal behavior.
At inference time, the model receives a prompt, which may include a system prompt with standing instructions, retrieved documents, and the user’s message. It generates the answer one token at a time, each token conditioned on everything before it. The prompt and the answer together must fit in the model’s context window. The model has no memory between requests unless the application supplies prior messages or stored facts.
Types of large language models
| Type | What it is | Typical business use |
|---|---|---|
| Base model | Pre-trained only; completes text but does not follow instructions reliably | Starting point for fine-tuning |
| Instruction-tuned model | Trained to answer requests and hold a conversation | Assistants, drafting, summarization, extraction |
| Reasoning model | Works through intermediate steps before answering | Multi-step analysis, code, and planning tasks that tolerate slower answers |
| Multimodal model | Accepts images, audio, or scanned documents as well as text | Reading forms, inspecting photos, summarizing meetings |
A second split cuts across all four types. Closed-source models are reached only through a vendor’s API. Open-weight models can be downloaded and run on infrastructure the company controls, which matters for data handling and for cost at steady volume.
Large language models in practice
A regional distributor receives several hundred purchase orders a day as PDFs and emails in inconsistent layouts. An LLM reads each one, extracts the customer, line items, quantities, and requested dates into a fixed structure, and flags anything that does not match the price list. A clerk reviews the flagged orders and the rest flow into the ERP system. The model replaced a manual keying step that occupied two people most of every morning.
A law firm uses a model with a long context window to compare a counterparty’s draft against the firm’s standard positions and list every deviation with its clause reference. The output is a starting point for the associate, who still reads the clauses that matter.
Large language models and private AI
Where the model runs decides who sees the data. Sending prompts to a hosted API means customer records, contracts, or patient notes leave the company’s environment under the vendor’s terms. Running an open-weight model on the company’s own GPUs or in a dedicated cloud tenancy keeps that data inside the boundary and turns a per-token bill into a fixed capacity cost. Kirk Tech Solutions typically recommends owning the model where data is sensitive or volume is steady, and renting through an API for experiments and low-volume tasks.
What to watch for
- LLMs produce fluent text that can be wrong, so any output that drives a decision needs grounding in retrieved documents or a human check.
- A model knows nothing after its knowledge cutoff and nothing about the company’s data unless the application supplies it.
- Benchmark scores on public tests do not predict performance on a firm’s own documents, so test with real samples before choosing a model.
- Count the full cost of running the model, including context length, concurrency, and the GPU memory a quantized version needs.
- Ask any vendor whether prompts are logged, retained, or used for training, and get the answer into the contract.