An AI model is a file of learned numbers, called parameters, plus the program that runs them. It turns an input into an output, such as a prediction, a label or generated text. The model bases that output on patterns it learned from training data.
Every AI model has two phases. It learns once during training, and then it is used many times during inference. Inference is the moment a trained model takes a new input and produces an answer.
In McKinsey's 2025 State of AI survey, 88 percent of respondents report regular AI use in at least one business function. A year earlier, that figure was 78 percent.
With that much use, the word "model" now shows up in many settings. This article covers AI models as software, not the AI-generated influencers that are sometimes also called "AI models."
The Three Parts of an AI Model: Architecture, Parameters, and Training Data
Every AI model is built from three parts. The architecture sets the structure, the parameters store what was learned, and the training data supplies the examples.
- Architecture: The shape of the math, such as a neural network, which passes numbers through layers of connected nodes.
- Parameters: Adjustable numbers, also called weights, that training tunes until the model's outputs match its examples well.
- Training data: The text, images or other examples the model learned its patterns from.
A deep learning model is a neural network with many layers. A 2025 survey of large language models reports that most modern language models build on the transformer, a deep learning design introduced in 2017. The transformer is good at tracking how each word in a passage relates to the others.
Parameter counts are easier to picture with a real example. Meta's Llama 3.1 announcement lists versions with 8 billion, 70 billion and 405 billion parameters. Meta trained the 405B model on over 15 trillion tokens using over 16 thousand H100 GPUs.
A token is a small chunk of text, such as a short word or part of a longer word. Models read and write text as tokens, so token counts measure both training data and the length of an answer.
Parameter count roughly signals a model's capacity to store patterns, but it does not measure quality. Data quality, training methods and fine-tuning also shape results.
Public parameter counts are also getting harder to find. Stanford's 2026 AI Index says parameter counts have stayed near 1 trillion for three years, but frontier labs no longer report them.
How AI Models Learn: Training vs. Inference
Training adjusts a model's parameters to reduce its errors on examples, and inference uses those fixed parameters to answer new inputs. Training happens rarely and costs a lot of computing power. Inference happens on every single request.
During training, the model makes a guess, checks the guess against the correct answer and nudges its parameters to shrink the error. It repeats this loop many times over. For a deeper walkthrough, see how AI model training works.
| Attribute | Training | Inference |
|---|---|---|
| What happens | The model learns patterns from examples | The model answers a new input |
| Parameters | Change with every update | Stay fixed |
| How often | Once per model version, plus occasional fine-tuning | On every request |
| Cost | Large GPU clusters and long training runs | A small amount of compute per request |
| Output | A trained model file | A prediction, label or generated text |
| Who does it | The lab or team that builds the model | The app, developer or user who runs the model |
The Main Training Steps
Most machine learning models follow the same basic training sequence:
- Collect and clean the training data.
- Train the model by adjusting its parameters to reduce errors.
- Evaluate the model on held-out data it never saw during training.
- Fine-tune the model on a smaller, focused dataset for a specific task.
Models also differ in how they learn from data. Three learning styles cover most cases:
- Supervised learning: The model learns from labeled examples, such as emails marked "spam" or "not spam."
- Unsupervised learning: The model finds groups and patterns in data that has no labels.
- Reinforcement learning: The model learns by trial and error, earning rewards for good actions and penalties for bad ones.
Held-out data catches a problem called overfitting, where a model memorizes its training examples instead of learning patterns that work on new inputs. To build one yourself, follow this step-by-step guide to training a model.
Why a Model's Knowledge Has a Cutoff
A model's parameters freeze when training ends, so it knows nothing that happened after its training data was collected. That last date is called the knowledge cutoff. Ask a model about events after its cutoff, and it has no stored facts to draw from.
The cutoff explains why a chatbot can sound out of date. A model released this year may still describe last year's prices, product versions or news as current. In its guide to Anthropic's model knowledge cutoffs, Anthropic warns that its models "may not be aware of events or information that occurred after their respective cutoff dates."
Types of AI Models
The main types of AI models differ by what they output. Some predict a label or number, while others generate new content. The categories overlap: a large language model is a generative model, and many are also foundation models.
| Type | What it does | Typical output | Example |
|---|---|---|---|
| Predictive (discriminative) model | Sorts or scores an input | A label, score or number | A spam filter that marks emails |
| Generative model | Creates new content that resembles its training data | Text, images, audio or code | An image generator that draws from a text prompt |
| Large language model (LLM) | Predicts the next token in a sequence of text | Written answers, summaries or code | GPT, Claude or Llama |
| Foundation model | Learns general patterns from broad data, then gets adapted to many tasks | A base model that teams fine-tune | Llama 3.1, which developers fine-tune for their own tasks |
| Computer vision model | Reads and interprets images or video | Object labels, boxes or captions | A model that spots defects in factory photos |
| Multimodal model | Handles more than one kind of input, such as text and images | Text that describes or combines several inputs | Gemini answering a question about a photo |
A large language model learns from huge amounts of text and writes by predicting one token at a time. Well-known LLM families include GPT from OpenAI, Claude from Anthropic, Gemini from Google and Llama from Meta. Mistral, from the French company Mistral AI, and DeepSeek, from the Chinese company of the same name, also publish widely used models.
AI Model vs. Algorithm vs. AI App vs. AI Agent
An algorithm is the learning recipe, and the model is what that recipe produces. An app wraps one or more models with an interface and extra features. An agent uses a model to plan steps and call tools in a loop until it finishes a task.
| Term | What it is | What it does | Example |
|---|---|---|---|
| Algorithm | A set of rules for learning from data | Adjusts parameters during training | Gradient descent, which nudges weights to cut errors |
| AI model | The trained file of parameters plus code that runs it | Turns an input into an output | A GPT model or Llama 3.1 405B |
| AI app | A product built around one or more models | Adds an interface, memory, search and safety filters | ChatGPT |
| AI agent | A system that uses a model to decide its next action | Plans, calls tools and checks results in a loop | An assistant that searches the web, reads pages and drafts a report |
ChatGPT is not an AI model. It is an app built on OpenAI's models, such as GPT-5, with chat history, file uploads, web search and safety filters added around them.
At its August 2025 launch, OpenAI said in Introducing GPT-5 (OpenAI) that GPT-5 "is the new default in ChatGPT." When people say "the model said," they usually mean the whole app.
The NIST AI Risk Management Framework draws a similar line by defining an AI system, not a model. NIST describes an AI system as a machine-based system that can "generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments." The same framework warns that training data can become stale compared with the setting where a system is deployed.
The model is one part inside that larger system. Agents add one more layer on top of the model. Read more about what AI agents are and how they use tools.
How AI Models Get Fresh Information at Inference Time
AI products get current facts by adding fresh data to the prompt at inference time, because the model's parameters stay frozen. The model then reads that data the same way it reads your question. Nothing about the model itself changes.
This step is called grounding: tying a model's answer to specific source material supplied with the request. Two common methods are retrieval-augmented generation and live web search.
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation (RAG) fetches relevant documents, places them in the model's context window and asks the model to answer from them. The documents can come from a company's files, a database or the web.
The context window is the amount of text a model can read at once, counted in tokens. It holds your question, any retrieved documents and the model's answer. If the documents don't fit, the app has to trim them or pick the most relevant parts.
Live Web Search and Tools
Models can also call search or scrape tools to read current web pages and return the sources they used. The app runs the tool, passes the page text into the context window and the model writes an answer from it. This is how a chatbot with an old knowledge cutoff can still answer questions about last week.
Web access can reduce hallucination, which is fluent but false output. The OpenAI GPT-5 system card shows the factual error rate of OpenAI's o3 model on a biography benchmark (FActScore) falling from 24.2% with browsing off to 5.7% with browsing on. OpenAI ran these tests itself, and the gains varied by model and test.
Source links let readers check each claim against the page it came from. That practice is the core of grounding LLM answers in sources. Without sources, a reader has no easy way to tell a grounded answer from a guess.
Olostep's Answers API handles this search, read and answer step in one API call. It searches the web, reads relevant pages and synthesizes an answer, and the response includes a sources array listing the URLs used.
For builders, the practical rule is to give the model clean input. Convert pages to Markdown or JSON before they reach the context window, and keep each source URL attached so answers can be verified.
Where AI Models Fail: Hallucination, Staleness, and Bias
AI models fail in three common ways: they invent facts, they go out of date and they repeat patterns from skewed data. Each failure traces back to how training and inference work.
- Hallucination: The model produces fluent but false output, because it predicts likely text instead of looking up facts.
- Staleness: The model's knowledge ends at its training cutoff, so recent events are missing.
- Bias: The model repeats unfair or skewed patterns that were present in its training data.
Hallucination rates differ widely between models. On Vectara's hallucination leaderboard, rates ranged from 1.8% to 24.2% across the models tested, as of October 2026. Vectara, a company that sells RAG software, runs this benchmark, which measures only how faithfully models summarize short documents.
Retrieval reduces hallucination but does not remove it. In February 2025 testing, the Columbia Tow Center AI search study asked eight AI search tools to identify the source of news excerpts. The tools "provided incorrect answers to more than 60 percent of queries."
Learn more about why AI models hallucinate and how to spot it. Builders can lower the risk by checking answers against their sources and testing the model on real user questions.
Frequently Asked Questions
Most beginner questions about AI models come down to one split: the model itself versus the product built around it. These answers keep that split in view.
Is ChatGPT an AI Model?
No, ChatGPT is an app built on OpenAI's models, such as GPT-5. The app adds chat history, web search and safety filters on top of the models.
What Are the Most Well-Known AI Models?
The best-known model families are GPT from OpenAI, Claude from Anthropic, Gemini from Google, Llama from Meta, Mistral and DeepSeek. Each family includes several versions of different sizes.
Is an LLM the Same as an AI Model?
An LLM is one type of AI model, built to read and write text. Many AI models are not LLMs, such as spam filters and image recognition models.
What Is the Difference Between an AI Model and an Algorithm?
An algorithm is the procedure used to learn from data, and the model is the trained result. Engineers can run the same algorithm on different data and get different models.
Can an AI Model Learn From My Conversations?
A model's parameters do not change while you chat with it. An app may save your conversations, and the provider may use them to train future models, depending on its policies and your settings.
Why Does an AI Model Have a Knowledge Cutoff?
A model only knows what was in its training data, and its parameters freeze when training ends. Apps work around the cutoff by adding search results or documents to the prompt at inference time.