What is GPT? — Its meaning and how to use it yourself
is not a . It is a trained to predict the next . Learn what people mean by GPT and how to use it yourself.
GPT is a model, not a program
The core of GPT is a huge set of numbers learned in (the ) and the way to calculate with them (the structure). So it cannot run by itself. It needs a separate that loads these numbers and does the calculation.
Think of music: the is the MP3 file, the program that runs it is the music player, and is the whole music app.
-
Generative
It does not look up stored text. It makes new , one at a time.
-
Pre-trained
Before you use it, it has already practised guessing the next token on a lot of text.
-
Transformer
A design published in 2017 that compares all the input tokens with each other.
What people mean when they say GPT
-
The design
The structure that sets the order of the calculations. This is the closest thing to "".
- In music terms
- The rules of the MP3 format
- When people say
- "A small model built with the GPT structure"
-
The trained model
The set of numbers you get by training that design. This is what GPT usually means.
- In music terms
- An MP3 file
- When people say
- "GPT-2 is public"
-
The program that runs it
The that loads the model and calculates. It is not the model, but people often say both as one.
- In music terms
- A music player
- When people say
- "I run GPT on my computer"
-
Software like ChatGPT
A product that adds a chat screen, accounts and other features to the model. In everyday talk, GPT often means this.
- In music terms
- The whole music app
- When people say
- "I asked GPT"
How is GPT different from an LLM?
- (large language model)
- LLMs built with a transformer
- The family
- LLMs built with a transformer
LLM is the name for every large language model trained on a lot of text, and GPT is the name of one family of them. It is like "car" and the name of one car model. An LLM is also a model made of a set of numbers and a calculating structure, just like GPT.
GPT is an LLM.
Not every LLM is a GPT. , and are LLMs, but they are not called GPT.
Are Claude Opus and DeepSeek GPT too?
By name they are not GPT, but by the way they are built they belong to the same family. In everyday talk, "not GPT" is the right thing to say.
-
Product or model name
No
GPT is usually the name of 's model family, such as GPT-4 and gpt-oss. Claude is 's model and is DeepSeek's.
-
LLM (large language model)
Yes
All three are models trained on a lot of text to predict the next token.
-
What the letters G, P and T mean
Mostly fits
They make new text (Generative), are trained beforehand (Pre-trained) and use a transformer structure (Transformer).
Think of cars: LLM is the word "car", and GPT, Claude and DeepSeek are model names from different companies. A car that works in a similar way is still not called by another company's model name.
What is not certain
- DeepSeek: its papers and public weights let you check for yourself that it uses a transformer structure. More about DeepSeek
- Claude Opus: Anthropic has not published its structure in detail. It is widely known to be in the transformer family, but you cannot check the details from public material the way you can with DeepSeek.
Is it used in development, or for making tokens?
GPT is less a development tool than the result of development. Once finished, it is used to make tokens.
-
Training Adjusts the numbers by practising next-token guesses
-
GPT model The result of training
-
The finished model makes tokens
You can also call a ready-made GPT through an to build a translator or a . Then GPT is a part you use in development.
The four steps that make a token
-
Split the text into tokens
-
Calculate the odds of the next token
-
Pick one token
-
Add it to the text
Repeat from step 2 until done
The program that runs all four steps is the runner. The model's numbers are used mainly for the calculation in step 2.
Names you often hear
- A runner for models in GGUF files
- A tool that downloads and runs a model with one command
- Not a runner: a file format for storing models
- A runner for ONNX files
- ChatGPT and the API
- The model runs on OpenAI's servers, not on your computer
Four ways to use GPT yourself
-
Use ChatGPT
Open the website or app and start talking. No needed.
- You need
- A web or the app
- Good for
- Trying it for the first time, daily life and work
-
Call the API from your program
Your program sends a question over the internet and gets the answer back.
- You need
- An , and a fee for the tokens you use
- Good for
- Putting it in your own service or app
The shortest first request curl https://api.openai.com/v1/responses \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $OPENAI_API_KEY" \ -d '{ "model": "gpt-6-astra", "input": "Hello!" }' -
Run it on your computer
Download and run a model with public weights, such as gpt-oss (). The easiest tool is .
- You need
- For gpt-oss:20b: a 14GB download and at least 16GB of
- Good for
- Keeping your data on your own machine
After installing Ollama, in a terminal ollama run gpt-oss:20b -
Train a small GPT yourself
Build a small GPT in code and train it on text. This is the way to learn how it works. A small model will not perform like ChatGPT.
- You need
- A browser (), or and
- Good for
- Learning how GPT works
Build it yourself with XSLM Build a GPT on your computer: GPT-2 in miniature
Try ChatGPT first, use the API to put it in your program, and train a small GPT if you want to know how it works.
How GPT got here
-
GPT-1 showed that a pre-trained transformer helps with understanding text.
-
: the largest model has 1,500,000,000 . It was released in stages, with the full weights out that November.
-
GPT-3 has 175,000,000,000 parameters. Its paper showed it doing new tasks from only a few examples.
-
ChatGPT: a chat service based on GPT-3.5 was released, and the name GPT became widely known.
-
GPT-4 takes pictures as input as well as text.
-
gpt-oss: models with weights anyone can download were released under the Apache 2.0 .
-
Models recommended by the OpenAI API docs:
gpt-6-astragpt-6.1-solgpt-6-luna
Watch out
-
The gpt of disks is a different word
The gpt in commands like parted is GUID Partition Table, a way of laying out a disk. It has nothing to do with this model.
-
Answers can be wrong
It chains tokens by probability, so it can write text that sounds right but is wrong. Check important things against the original source.
-
An API key is a secret
Do not write it in code you push to a repository or put it in a web page. Keep it in an environment variable, as in the example.
-
Model names change often
The model names on this page are as of 2026-10-31. Check the official model list before you use one.
Sources
This page is a plain-language summary of the sources below, checked on 2026-10-31. If they differ, follow the source.
- OpenAI API Docs — Models(opens in a new tab)
- OpenAI API Docs — Quickstart(opens in a new tab)
- OpenAI — Improving language understanding with unsupervised learning (GPT-1)(opens in a new tab)
- OpenAI — Better language models and their implications (GPT-2)(opens in a new tab)
- OpenAI — GPT-2: 1.5B release(opens in a new tab)
- Language Models are Few-Shot Learners (GPT-3, arXiv)(opens in a new tab)
- OpenAI — Introducing ChatGPT(opens in a new tab)
- OpenAI — GPT-4(opens in a new tab)
- OpenAI — Introducing gpt-oss(opens in a new tab)
- DeepSeek-V3 Technical Report (arXiv)(opens in a new tab)
- Ollama — gpt-oss(opens in a new tab)
- Attention Is All You Need (Transformer, arXiv)(opens in a new tab)