Model Card

A model card is a structured document published by an AI provider describing a model's intended use, training data, performance, limitations, and ethical considerations.

What is a model card?

A model card is a structured documentation artifact published by an AI provider describing a specific model: its intended use cases, training data sources and processing, evaluation metrics across demographic and use-case slices, known limitations, and ethical considerations. The format was proposed in the 2018 paper "Model Cards for Model Reporting" by Mitchell et al. and has become a de facto industry norm — Hugging Face, Meta, Google, and Microsoft all publish model cards for their open-weight releases.

What buyers should look for

A high-quality model card answers: what task is this model designed for, what training data was used and how was it filtered, what is the model's performance broken down by demographic group / language / domain, what known failure modes exist, what is the recommended use vs out-of-scope use, and who should be contacted with concerns. The absence of any of these in a model card is itself information — many proprietary frontier models publish minimal cards, citing competitive concerns.

Model cards in regulated procurement

The EU AI Act references documentation requirements aligned with the model card concept. NIST AI RMF's Generative AI Profile recommends model cards as part of the Map function. For regulated buyers (healthcare, finance, government), the absence of a model card from an AI vendor is a meaningful signal — and an opportunity to ask the vendor for an equivalent document under NDA if their public posture is intentionally sparse.