A large language model (LLM) is an AI system trained on vast amounts of text to predict the next unit of language and, from that, to read, summarize, and answer questions about documents. In commercial real estate extraction, an LLM reads a lease, rent roll, or offering memorandum and returns structured fields such as base rent or term.
How Does a Large Language Model Work?
A large language model works by converting text into tokens, then predicting the next token from the sequence before it, one token at a time. A token is roughly four characters of English, so 100 tokens is about 75 words, per OpenAI's tokenizer documentation. The model learns these predictions from billions of parameters tuned during training on large text corpora.
The model reads only what fits in its context window, the maximum number of tokens it can hold at once. GPT-4 launched with an 8,192-token window, later raised to 128,000 tokens, and GPT-4.1 reached 1 million tokens in April 2025, per the GPT-4 Wikipedia entry and OpenAI documentation. Anthropic's Claude models default to a 200,000-token window, which Anthropic describes as roughly 150,000 words or over 500 pages.
Concept | What it means in extraction |
Token | About four characters of English text, the unit the model reads |
Parameter | A learned weight; GPT-3 had 175 billion, per NVIDIA |
Context window | Maximum tokens the model can read at once, such as 128,000 |
Prompt | The instruction plus document text handed to the model |
Why a Large Language Model Matters
A large language model matters in commercial real estate because it reads unstructured documents the way a person does, not the way a rules engine does. A lease phrases the same clause a hundred ways across a hundred landlords. An LLM generalizes across that variation, so it can find base rent whether the lease calls it "Base Rent," "Minimum Annual Rent," or "Fixed Rent."
The context window sets a hard limit worth respecting. At roughly 0.75 words per token, a 128,000-token window holds about 96,000 words, near 192 pages at 500 words per page. A large offering memorandum or a full lease with exhibits can exceed that, which is why long documents are chunked or retrieved rather than pasted whole. An LLM that cannot see a clause cannot extract it.
Example
A large language model reads a 90-page retail lease to pull annual base rent. At about 500 words per page, the lease is roughly 45,000 words. Dividing 45,000 words by 0.75 words per token gives about 60,000 tokens, which fits inside a 128,000-token context window with room for the prompt and the answer.
Step | Figure |
Lease length | 90 pages at 500 words per page = 45,000 words |
Token estimate | 45,000 words / 0.75 words per token = 60,000 tokens |
Context window | 128,000 tokens |
Headroom | 128,000 - 60,000 = 68,000 tokens for prompt and output |
Because the whole lease fits, the model reads all 90 pages in one pass and returns the base rent figure of $612,000 with the governing schedule identified. Had the lease run 300 pages, roughly 200,000 tokens, it would exceed the window and would need chunking or retrieval so the model saw the rent schedule at all.
Variations and Edge Cases
Large language models vary by size, training, and how they are pointed at a document. The choices below change accuracy, cost, and whether a long lease fits in one pass.
Variant | Treatment |
Base model | Predicts text but is not tuned to follow instructions |
Instruction-tuned model | Fine-tuned to follow prompts, the usual choice for extraction |
Small vs frontier model | Smaller models cost less but miss subtle clauses |
Long-context model | Windows of 1 million tokens fit an entire deal room |
Multimodal model | Reads scanned pages and tables as images, not text alone |
Large Language Model vs Machine Learning Model
A large language model is often confused with a machine learning model in general, and the difference is scope. A machine learning model is any system that learns a mapping from data, including small classifiers that predict one label. A large language model is one specific, very large kind of machine learning model, trained on text to predict tokens and generate language.
Every large language model is a machine learning model, but most machine learning models are not large language models. A model that flags a rent roll row as "vacant" from a few numeric features is a machine learning model. A model that reads the lease narrative and writes a summary is a large language model.
Frequently Asked Questions
What is a large language model in simple terms?A large language model is an AI system trained on huge amounts of text to predict the next word, which lets it read, summarize, and answer questions about documents. In commercial real estate it reads leases and offering memorandums and returns structured fields such as rent and term.
How much text can a large language model read at once?A large language model can read only what fits in its context window, measured in tokens. Common windows range from 128,000 tokens, about 192 pages, to 1 million tokens for the newest models, per OpenAI documentation. Documents longer than the window must be chunked or retrieved.
Is a large language model the same as ChatGPT?No. ChatGPT is a product built on large language models such as the GPT series. The large language model is the underlying system that predicts text; the chat product wraps it with an interface, guardrails, and tools.
Related Terms
Machine Learning Model
Training Data
Retrieval-Augmented Generation
AI Hallucination
Document Extraction