Base vs instruct model

A base model is the raw result of pretraining and continues whatever text it is given; an instruct (or chat) model has been post-trained to follow instructions in a conversation format.

Why it matters. Most applications want the instruct model; base models suit fine-tuning, research and text scoring.

How it works

Given "Write a haiku about rain", a base model may continue with more prompts or a list, because it predicts likely text. An instruct model, trained on conversations in a chat template, answers the request. Reasoning variants are post-trained further to think before answering.

Open-weights makers often release both, with suffixes such as Base, Instruct or it; API models are almost always instruct models. Base models usually have no safety training and no chat template, and their benchmark scores are typically reported with few-shot prompts, so they are not directly comparable with instruct scores.

Use instruct models in applications. Fine-tune a base model when you want full control over format and behaviour and have enough data; with a small dataset, fine-tuning an instruct model is usually easier.

Worked example

Illustrative: prompted with "Q: What is the capital of France? A:", a base model may write "Paris" and then carry on with "Q: What is the capital of Spain?" until it hits a 50-token limit, paying for about 50 tokens. An instruct model, given the question in its chat template, answers "Paris." in a few tokens and stops.

Common mistakes

Deploying a base checkpoint in a chat app because its benchmark scores looked higher.
Fine-tuning a base model on a few hundred examples and expecting instruct-level behaviour.
Sending plain text to a self-hosted instruct model without its chat template.