Generative AI Leader Practice Question: The fundamental difference between a foundation…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  A foundation model is pre-trained on a large, diverse corpus; a fine-tuned model is adapted from a foundation model on a specific domain or task

  Why this is correct

  The axis of difference is training provenance: a foundation model is pre-trained from scratch on a broad, diverse corpus, whereas a fine-tuned model starts from that foundation and is further adapted on domain- or task-specific data, shifting its behaviour toward the target use case.

• ✗

  Foundation models only generate text; fine-tuned models can generate images

  Why it's wrong here

  Modality is set by the model architecture and training data, not by whether it is a foundation or fine-tuned model; fine-tuned image and multimodal models exist alongside text-only foundation models. The option tempts because early foundation models were text-focused, but the actual distinction is general pre-training versus task-specific fine-tuning.

• ✗

  Foundation models are open-source; fine-tuned models are proprietary

  Why it's wrong here

  Licensing is independent of the foundation-versus-fine-tuned distinction: open-weight foundation models exist, and fine-tuned derivatives can remain proprietary or be released openly. The option tempts because many well-known foundation models are open-weight, but the fundamental difference is broad pre-training versus adaptation to a specific task or dataset.

• ✗

  Foundation models require no inference infrastructure; fine-tuned models do

  Why it's wrong here

  Both foundation and fine-tuned models run inference on the same serving infrastructure; fine-tuning changes weights, not whether inference is required. The option is tempting because fine-tuning adds training compute, but that is a training-time cost, not an inference-infrastructure split. The real axis is pre-training versus task-specific adaptation.

Go deeper

Related to this question