The one-sentence version
Size (previous pages) is how much a model knows. Type is what the model is shaped to do. Picking the right type matters as much as picking the right size.
Chat (Instruct) Models
The default — and what you've been using all along.
A chat model (often labeled instruct) is trained to follow instructions and converse: you type a request, it answers. Everything in the main Academy — delegation, the Playbook, the recipes — assumes this type.
Use for: everyday delegation — drafting, summarizing, Q&A, email, working through tasks with a co-worker. This is the right default 90% of the time.
Reasoning Models
Think first, answer second — visibly.
A reasoning model produces a hidden (or visible) chain of thought before its final answer: it works the problem step by step, checks its own logic, then answers. That extra thinking takes extra time and tokens — but on hard problems it's the difference between a guess and a solution.
Use for: multi-step analysis, math, code debugging, planning, tricky judgment calls — anywhere the path to the answer matters.
Don't use for: quick lookups, simple rewrites, chat. Paying thinking-time tax on easy tasks just makes them slow.
Embedding Models
The quiet workhorse behind search and RAG.
An embedding model doesn't write text at all. It reads text and outputs a list of numbers — a vector — that encodes the meaning. Two passages with similar meaning get similar vectors, which makes "find me the document that means the same thing as this question" a math problem instead of a keyword problem.
Where you've met it: every time an agent "reads your documents and answers questions," an embedding model did the finding and a chat model did the answering. That pairing is RAG (see the Glossary).
Embedding models are tiny and fast — usually well under 1B parameters.
Vision & Multimodal Models
Models that read more than text.
A vision-language model accepts images alongside text: photographs, screenshots, scans, charts. A fully multimodal model handles several media — text, images, sometimes audio — in one model. Business superpower: feeding an invoice scan or a dashboard screenshot straight into the workflow.
Use for: reading scanned documents, extracting data from forms and screenshots, describing or comparing images.
Mixture of Experts (MoE)
Big total, small working set.
A mixture-of-experts model contains many "expert" blocks but routes each token through only a few of them. You'll see names like A3B (active) vs total parameters: a "402B-A38B" model has 402B parameters inside but uses only ~38B per token. Result: frontier-class knowledge with much lower running cost than a dense model of the same size.
Watch the memory, though: all 402B parameters still need to sit in storage/memory even if only some fire per token. MoE trades compute cost, not capacity cost.
Coder Models
Specialists tuned on code.
A coder model is a general model fine-tuned heavily on programming languages. At equal size it writes better code than its general sibling — and worse poetry. Many teams run one small general chat model plus one coder model side by side.
Use for: writing and reviewing code, SQL, regular expressions, and automation scripts — including n8n workflow code.
Quick Reference
| Type | Best at | Typical size |
|---|---|---|
| Chat / Instruct | General work with your AI co-worker | 1B–400B+ |
| Reasoning | Hard, multi-step problems | 7B–700B+ |
| Embedding | Search & retrieval (feeds RAG) | 0.1–1B |
| Vision / Multimodal | Images, scans, screenshots | 4B–100B+ |
| MoE | Big quality, cheaper tokens | 30B–700B total |
| Coder | Code & technical work | 7B–100B+ |