A model trained to produce new outputs such as text, images, or audio rather than only assign labels to fixed categories.
What It Is
Generative models optimize objectives that support sampling or decoding: next-token prediction, denoising score matching, or latent diffusion. They may start from a foundation model base but the generative role describes what the trained artifact does at inference, not its parameter count alone.
Why It Matters
Separating generative from discriminative models clarifies evaluation (likelihood, sample quality) and which atlas families—diffusion, autoregressive language models—apply to a product.
Simple Example
An autoregressive language model that completes prompts token by token is generative. A spam filter that outputs spam versus not-spam without synthesizing new mail is discriminative, even if both use transformers.
Common Confusions
Generative versus discriminative is about the training objective and inference behavior, not model size. A generative model is not the same as a foundation model: many foundation checkpoints are adapted into discriminative classifiers. Generative models also differ from representation learning alone—representations can support both roles depending on the head.