An AI model is a program that has learned patterns from data and uses them to make predictions or create new content. A language model predicts the next word. An image model predicts pixels that match a description.
The model is the trained engine. The app is what you use to reach it. GPT, Claude and Gemini are families of models. ChatGPT, the Claude app and Sparkian are apps that run them.
How an AI model is made
- The model learns from a large set of examples, such as text, images or audio. This is called training data.
- Training, tt adjusts billions of internal numbers, called parameters, until its predictions improve.
- Extra training on instructions and feedback makes it follow requests and behave safely. This is called Tuning.
- You send a prompt and the finished model produces an output. This is the part you pay for in an app. This is called Inference.
Training is done once per version. A model doesn’t learn from your chats as you use it. Features like AI memory store notes outside the model.
Types of AI models
| Type | Input to output | Examples of use |
|---|---|---|
| Language model (LLM) | Text to text | Chat, writing, code |
| Multimodal model | Text and images to text | Reading charts, screenshots, PDFs |
| Image model | Text to image | Illustrations, product shots |
| Speech model | Audio to text, or text to audio | Transcripts, voiceovers |
| Video model | Text or image to video | Short clips |
| Embedding model | Text to numbers | Search and knowledge bases |
Some language models are reasoning models that think before they answer. Some are open-source LLMs with public weights you can run yourself.
Model names and versions
Model names usually combine a family, a version and a size or tier. In “Claude Haiku 5.5”, Claude is the family, Haiku is the fast tier and 5.5 is the version.
Larger tiers tend to be smarter, slower and more expensive. Smaller tiers are faster and cheaper. New versions arrive every few months, so a model that led last year may now trail.
Choosing a model in Sparkian
Sparkian lists 100+ models from OpenAI, Anthropic, Google, DeepSeek and others in one picker. I can filter by provider or by use, such as General, Reasoning or Coding.

The details panel helps me match a model to the job. You can read more about each family on the AI models page.
