What is an instruct model?
An instruct model (also called an instruction-tuned or chat model) is a base large language model that has been fine-tuned to follow human instructions. A raw base model only knows how to predict the next token, so if you type a question it might continue the question or produce a list of similar questions. An instruct model, by contrast, interprets your prompt as a request and tries to fulfil it. Every consumer AI assistant — ChatGPT, Claude, Gemini — is built on an instruct variant rather than a raw base model.
Base model vs instruct model
The distinction matters in practice:
- Base model — trained purely to predict the next token across vast text. Ask it “What is the capital of France?” and it may answer, continue the sentence, or generate more questions. Useful for fine-tuning, completion, and research.
- Instruct model — further trained so that a prompt is treated as an instruction. Ask the same question and it directly replies “Paris.” This is the behaviour people expect from an assistant.
How instruct tuning works
Turning a base model into an instruct model usually involves two stages:
- Supervised fine-tuning (SFT) — the model is trained on a curated dataset of instruction-and-ideal-response pairs, teaching it the format of being helpful.
- Reinforcement learning from human feedback (RLHF) — humans (or a reward model trained on human preferences) rank multiple responses, and the model is optimised to prefer answers people rate as more helpful, harmless, and honest.
Some labs add variations such as Constitutional AI or direct preference optimisation (DPO), but the goal is the same: align the model’s behaviour with what users actually want.
How to tell them apart
Two quick signals:
- Naming — instruct variants often carry suffixes like
-instruct,-chat, or-it(for “instruction-tuned”). A name without such a suffix is frequently a base model. - Behaviour — give it a bare instruction. A base model tends to continue or imitate your text; an instruct model answers the request.
For building chatbots, assistants, and most product features, you almost always want the instruct version. Base models are mainly useful when you intend to fine-tune further or need raw completion behaviour.