Terms and model types
Language model, image model, embedding, agent, RAG: which kind of system solves which task and how to tell them apart.
The types you have to tell apart
| Type | Input | Output | Used for |
|---|---|---|---|
| Language model | Text | Text | Writing, summarising, classifying |
| Embedding model | Text or image | A vector | Search, similarity, grouping |
| Image model, recognising | Image | Label, boxes, mask | Checking, counting, sorting |
| Image model, generating | Text | Image | Illustration, drafts |
| Speech recognition | Audio | Text | Dictation, minutes |
| Speech synthesis | Text | Audio | Read-aloud, announcements |
| Multimodal model | Text and image | Text | Understanding documents |
Constructions, not models
Three terms describe not models but how models are assembled:
- RAG: search your own documents first, attach the passages to the question, then answer. That is where verifiability comes from.
- Agent: give the model tools and let it decide which to call. That is where capability to act, and risk, come from.
- Fine-tuning: retrain an existing model on your own examples. That is where form comes from, not knowledge.
Which construction for which problem
| Problem | Right answer |
|---|---|
| The model does not know our content | RAG |
| Answers must be verifiable | RAG with a citation duty |
| The format is wrong | Prompt first, then fine-tuning |
| The tone is wrong | Fine-tuning |
| Several systems are involved | Agent, with narrow permissions |
| It is too expensive or too slow | Smaller model, quantisation, batching |
| We may not release data | Local model or on-device processing |
The most common wrong decision is fine-tuning where retrieval was meant. Knowledge belongs in documents, not in weights: documents can be updated, deleted and cited; weights cannot.
Orders of magnitude
| Size | Parameters | Runs on | Used for |
|---|---|---|---|
| Small | 1 to 4 bn | Laptop, phone | Classification, simple extraction |
| Medium | 7 to 30 bn | One graphics card | The working range for most tasks |
| Large | 70 bn and up | Several cards or a service | Hard reasoning, broad knowledge |
Terms that get mixed up
| Term | What it means | What it does not mean |
|---|---|---|
| Foundation model | Broadly pre-trained, intended as a basis | Not automatically large or good |
| GPAI model | The AI Act's legal term for general-purpose models | Not identical to language model |
| Open weights | Parameters downloadable | Not open source; the licence may restrict |
| Fine-tuning | Retraining on your own examples | Not knowledge injection |
| Reasoning model | Produces intermediate steps before answering | The chain of thought is not a log |
| Context window | How much is visible at once | Not memory across sessions |
The second entry has immediate legal consequences: general-purpose models carry their own duties on technical documentation, copyright policy, and, above a training-compute threshold, additional requirements for systemic risk. See Duties by role.
A decision aid
- 01
Does the necessary knowledge exist in writing?
If yes: retrieval. If no: the preparatory work is writing it down, not training a model.
- 02
Must the answer be verifiable?
If yes: retrieval with a citation duty and a verification step. Nothing else carries.
- 03
May the system act?
If yes: an agent, but with per-tool permissions, a step limit and confirmation before irreversible actions.
- 04
May data leave the building?
If no: a local model or on-device processing, and the model size follows the hardware you have.
Which model type fits
FREE ACCOUNT
Model type decision sheet
A mapping from task to model type, including the cases where no model at all is the better answer.
Worksheet2 items
Related courses and sources
Elements of AI
A free introductory course from the University of Helsinki, no prior knowledge required. Around 30 hours, with a certificate.
If you want somewhere to go after our beginner path, this is the obvious next step. It overlaps our foundations in places but goes considerably deeper into how things work.