Understanding models · Chapter 1 of 8
Kinds of models
Chat, reasoning, multimodal, embedding, reranking, speech, image, guard and small models do different jobs, and one agent often uses several.
Goals
- Fit the first slot of the voice agent: turn the caller's voice into text.To do
- Fit the two slots that find the notes that answer the question.To do
- Fit the last three slots: write the reply, check it and say it.To do
- Pass the check.To do
Step 1 of 4: The read
Different models, different jobs
A model is a program that learned patterns from examples. Models learn different jobs. Each takes in some kind of input, such as text, sound or images, and gives out some kind of output. An agent often chains several models, each doing the job it does best.
Models that write
A chat model, also called an instruct model, follows instructions and writes a reply. A reasoning model works through a problem step by step before it answers: better on hard problems, but slower. Many new models do both, with a setting for how much they think. A multimodal model also reads images, sound or video.
Models that find and sort
An embedding model turns a text into a list of numbers, so texts with similar meanings get similar numbers. A search can then compare the numbers and find notes close in meaning to a question, quickly, even among thousands. A reranker reads the question together with each note it is given and puts the best first. That is more careful but slower, so it only sorts the few notes the search found.
Models that listen, speak, draw and check
Speech to text writes down what someone says. Text to speech reads text out loud. An image model makes a picture from a description. A guard model, also called a safety classifier, labels text or images as safe or unsafe, so an agent can stop a bad reply before anyone sees or hears it.
Small, closed and open
A model's size is counted in parameters: the internal values it adjusted in training, also called its weights. Small models have a few billion parameters or fewer, so they can run on a phone or a laptop. A closed model's weights are not published, and most closed models are used through an online service. An open-weights model can be downloaded and run on your own computer.
In the visual you fit out a voice agent for a library's phone line, one slot at a time. The library is made up; every example model you meet is real, as of 2026-11-03.
Step 2 of 4: Try it: Fit out a voice agent
Fit out a voice agent
Fit one slot at a time, from the top: read what the slot takes in and must give out, then pick the model card that matches.
Loading the visual…
Mode:SimulatedSimulated: the library, its notes and the call are made up, and no model runs in this page. The example models, their sizes and the one price shown are real, as of 2026-11-03.
Step 3 of 4: Check
Check
Next: Question 2 of 3
Step 4 of 4: What you can do now
What you can do now
- Name ten kinds of models and say what each takes in and gives out.
- Pick a kind of model for each step of a simple agent, and say why another kind would not fit.
- Give a closed and an open example of each kind, from a dated snapshot.