---
title: "Large Language Model"
url: "https://mahouse.world/glossary/large-language-model"
type: "glossary term"
topic: "Language models"
author: "Ihor Vlasov"
reviewed: "2026-09-20"
published: "2026-08-06"
---

# Large Language Model

A large language model (LLM) is a neural network — a transformer with billions to trillions of parameters — trained on vast quantities of text to predict the next token, which turns out to be enough for it to write, summarise, translate, answer questions, reason and produce code; post-training turns the predictor into an assistant. GPT-4, Claude, Gemini and Llama are LLMs; for a fintech they power assistants and document processing, with hallucination and governance as the constraints.

**Also known as:** LLM, LLMs, Language model, Frontier model, Chat model

## Key facts

- **Used at:** customer service and assistants, document extraction and summarisation, knowledge search (with retrieval), classification and routing, code generation, compliance and legal drafting, analyst and research support, and, as agents, multi-step tasks with tools; across the economy, the largest technology adoption since the smartphone
- **Binding:** the EU AI Act's obligations on general-purpose AI model providers (transparency, training-data summaries, copyright policy, and systemic-risk duties above 10²⁵ FLOPs of training compute) from August 2025, with obligations on deployers by use; data protection on the data sent to models; sector rules on decisions models influence; the labs' terms of service on what the model may be used for and whether the vendor trains on customer data
- **Typical scale and cost:** frontier models trained on tens of trillions of tokens at costs of hundreds of millions of dollars; API prices from cents to tens of dollars per million tokens; context windows from 128,000 to over a million tokens; open-weight models from 1 to 700 billion parameters runnable on hardware from a laptop to a data centre
- **Typical limits:** hallucination (fluent, confident, wrong), a knowledge cut-off, sensitivity to prompt wording, inconsistency between runs, prompt injection through content the model reads, and cost and latency at scale; the mitigations are retrieval, verification, structured outputs, guardrails, evaluation and human review
- **Parties:** the labs (OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and others), the clouds that host them, the open-weight community, the companies building on them, the regulators, and the customers who talk to them

## How a large language model works in practice

A merchant asks a payments company's assistant why a payout is delayed. The assistant's system prompt sets its role, rules and format; a retrieval step finds the merchant's payout record and the relevant knowledge-base articles; the model receives all of it as context and generates an answer token by token, each token predicted from everything before. It explains that a reserve was applied after a chargeback spike, cites the policy, and offers the steps to appeal — in the company's tone, in the merchant's language, in two seconds. Nothing was scripted; the model composed the answer from the context and its training. The same model, given a bank statement, extracts transactions into a table; given a contract, summarises the termination clauses; given a description, writes the code.

The engineering that makes this reliable is around the model, not in it. Retrieval grounds answers in the company's documents rather than the model's memory. Structured outputs and verification catch the figure the model invented. Guardrails and restricted tools keep it from sending money because an email told it to. Evaluation sets measure every change; monitoring watches production; human review covers the decisions that matter. The choice of model is an evaluation on the company's own tasks, not a benchmark, and the smallest model that passes is the one to run — often a fine-tuned open model for narrow tasks, a frontier model for the hard ones. For AI M&A, LLMs are the reason the labs are valued as they are, the reason licence-and-hire deals were structured around teams that can train them, and the reason an application company's moat is its data, its workflow and its distribution — never the model it rents.

For a fintech, an LLM is a general capability rented by the token and made reliable by engineering; for a buyer, the target's value is in what it built around the model and what data it has, because the model is the same one its competitors can rent.

## Example

A lender's operations team spends 40 minutes per application on document review — bank statements, payslips, contracts — and 20 minutes on the file summary. It deploys an LLM pipeline: a vision-capable model extracts figures with citations to page and line; a verification step recomputes totals; a summarisation prompt produces the file note in the underwriter's format; an evaluation set of 500 applications is run on every model or prompt change. Time per application falls to 12 minutes, with the underwriter reviewing extracted figures against the cited sources and making the decision — the model recommends nothing. Cost: $0.60 per application in tokens. The model-risk committee classifies the pipeline as an extraction tool, not a decision system, on the strength of the design; the vendor's enterprise terms exclude training on the lender's data; and the pipeline's evaluation results are in the file when the regulator asks how AI is used in underwriting.

## Common mistakes

- Trusting an answer because it is fluent; verify what matters.
- Sending customer data to a model under consumer terms that allow training on it.
- Letting the model decide what a person must decide, and calling it assistance.
- Choosing a model by benchmark rather than by an evaluation on the company's own tasks.
- Valuing an application company on the model it rents rather than on its data and workflow.

## Related terms

- [Transformer](https://mahouse.world/glossary/transformer.md)
- [Pre-Training](https://mahouse.world/glossary/pre-training.md)
- [Instruction Tuning](https://mahouse.world/glossary/instruction-tuning.md)
- [RLHF](https://mahouse.world/glossary/rlhf.md)
- [Hallucination](https://mahouse.world/glossary/hallucination.md)
- [RAG](https://mahouse.world/glossary/rag.md)

— Ihor Vlasov, M&A FinTech Expert. Last reviewed 2026-09-20. https://mahouse.world/authors/ihor-vlasov