A model card is a short document published alongside an AI model that provides structured information about what the model does, how it was built, and where it should or should not be used. Standard sections cover the model's intended use cases, out-of-scope uses, training data sources, evaluation results across relevant benchmarks, known limitations and biases, and ethical considerations. The format was formalised in a 2019 Google Research paper and has since been adopted by Hugging Face as a community standard, by Anthropic and OpenAI for their own releases, and by regulatory frameworks that require AI transparency documentation.

Think of it like a nutrition label on packaged food: just as the label tells you what is in the product, how it was made, and who it is appropriate for, a model card tells you what a model is trained on, how it performs across groups, and which applications it is and is not suited for.

A minimal model card entry on Hugging Face is written in a YAML-headed markdown file:

---
language: en
license: apache-2.0
tags: [text-classification, sentiment]
datasets: [stanfordnlp/sst2]
metrics: [accuracy]
---
## Model Description
Fine-tuned BERT for binary sentiment classification on SST-2.
## Intended Uses and Limitations
Intended for English product reviews. Not validated for medical or legal text.

Model cards are closely tied to Fine-tuning (cards describe what base model was adapted and how), Bias documentation (a required section), RLHF (cards explain alignment procedures), and Foundation Model release practices.

You will encounter model cards whenever you browse Hugging Face model repositories, review an AI vendor's release documentation, or work in a regulated industry where auditors require evidence of responsible AI practices.