Instruct Model.

A base language model trained only to predict the next token will complete text in the style of its training corpus — it won't answer questions or follow directions in the way a user expects. An instruct model has been additionally trained (typically via supervised fine-tuning on human-written instruction-response pairs and RLHF) to be helpful, harmless, and honest.

The transformation from base model to instruct model is what makes GPT-4, Claude, and Gemini useful in practice. The same underlying architecture that generates arbitrary text is shaped by instruction tuning into a model that answers questions, writes code, and follows safety guidelines.

Many open-weight model families release both base and instruct variants. The base model (e.g. Llama 3.3 70B) is preferred by researchers who want to apply their own fine-tuning. The instruct model (e.g. Llama 3.3 70B Instruct) is ready to use out of the box for most applications.