Module 2: How a language model works · Lesson 2.6 · about 20 min
Base, instruct, reasoning and multimodal models
The same family often comes in several versions. What base, instruct (chat), reasoning (thinking) and multimodal models are, and which one to download.
Module 2 · about 20 minutes · read first: 2.3. You’ll be able to pick the right version of a model from a list of downloads.
1. Why this matters
Search for a model and you will see names ending in -base, -it, -instruct, -thinking, -vision. Download the wrong one and it will ramble, ignore instructions or refuse your image.
2. The versions
| Version | Names often include | What it does | Use it for |
|---|---|---|---|
| Base (pre-trained) | base, pt | Continues text; does not follow instructions well | Fine-tuning (Module 14), research |
| Instruct / chat | it, instruct, chat | Follows instructions, holds a conversation | Almost everything |
| Reasoning / thinking | thinking, reasoning, or a switch | Writes its working before answering | Maths, logic, planning, tricky code |
| Multimodal | vision, vl, omni | Takes images, sometimes audio | Photos, scans, screenshots, voice |
| Embedding | embed | Turns text into numbers for search | Document search (Module 12) |
Many recent models combine these: Gemma 4 instruct models take images, and several families let you switch thinking on or off per request (Lesson 10.4).
Here is how to read a typical download name, piece by piece. The pattern is common on Hugging Face, though every publisher varies it a little.
The file is called gemma-4-12b-it-Q4_K_M.gguf. Read it left to right.
gemma-4 is the family and generation: who made it and which release.
12b is the size: 12 billion parameters (Lesson 2.2).
it means instruction-tuned: the chat version, the one most people want. If it said pt (pre-trained) it would be the base model.
Q4_K_M is how tightly the numbers are squeezed, about 4 to 5 bits each (Module 8), and .gguf is the file format (Lesson 7.5). So: a chat model, about 7 GB, ready for llama.cpp, Ollama or LM Studio.
- base (pt)
- the model straight after its main training: it continues text but does not follow instructions well
- instruct (it, chat)
- the base model after extra training on instructions and good answers, plus feedback
- thinking
- trained to write out its working before the answer; slower, better at hard problems
- vision, vl, omni
- can read images (and sometimes audio) as well as text
- Physical meaning
- It is one education with different job training on top. Pick by the job: instruct for nearly everything, thinking for hard multi-step problems, vision for photos and scans, base only if you plan to train it further yourself.
“The base model is the pure, smarter version, and instruct is a watered-down copy.” They know about the same; the instruct version has extra training that teaches it to understand a request and answer it. Ask a base model to summarise an email and it may just write more emails, because “continue the text” is all it was trained to do.
Maths and principle
Principle: the same pre-trained weights are specialised by different post-training.
Equation: instruct model = base model + training on (instruction, good answer) pairs + feedback.
Analogy: a medical graduate (base) becomes a GP, a surgeon or a researcher depending on the residency (post-training). Same education underneath, different job.
Try it: you need to read text from photographed receipts and return JSON. Which two properties do you need? (Instruct and multimodal.)
Thinking has a price in time. A reasoning model writes its working as tokens before the answer, at the same speed as the answer itself.
3. Worked example
Made-up numbers. Arjun downloads a “-pt” file and asks it to summarise an email. It writes three more emails instead. He switches to the “-it” version of the same model and gets a summary first time.
4. Practice
- In the Ollama library or on Hugging Face, find a model with both base and instruct versions.
- Which version for a chatbot that answers questions about your documents?
- When would you turn thinking off?
Answers
- Most families publish both; Ollama’s default tags are usually instruct.
- Instruct, plus an embedding model for the search part.
- For quick, simple tasks where speed matters more: summaries, rewrites, short answers.
New words: base model, instruct model, reasoning model, multimodal, embedding model.
- Base (pt) models continue text; instruct (it, chat) models follow instructions and hold a conversation.
- Reasoning (thinking) models write out their working first: better on hard problems, slower and more tokens.
- Multimodal (vision, vl, omni) models take images and sometimes audio; embedding models turn text into numbers for search.
- Read the file name: family, size, version tag (it or pt), squeeze level (Q4_K_M) and format (.gguf).
- For almost every job, download the instruct version.
Check yourself
Sources
- Ouyang, L. et al. (2022), Training language models to follow instructions with human feedback (how base models become instruct models) https://arxiv.org/abs/2203.02155
- Wei, J. et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models https://arxiv.org/abs/2201.11903
- Hugging Face Hub documentation, model cards https://huggingface.co/docs/hub/model-cards
Your ticks stay in this browser. Save them as a file on the course home.
Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.