What it means
A model card is a short structured document accompanying a released model: its intended uses, out-of-scope uses, training data at a summary level, evaluation results including performance broken down across groups, and known limitations.
The idea came out of fairness research and has become common practice, with larger labs publishing extended system cards covering safety evaluations and red-teaming results for frontier releases.
Quality varies enormously. Some are genuinely informative about limitations; others are marketing with a documentation format. The most useful sections are consistently the least promoted ones — out-of-scope uses and disaggregated evaluation results, which is where a vendor states what the model is bad at.
Why it matters
Model cards are the main artifact for judging a model beyond its headline benchmark, and documentation of this kind is becoming a regulatory requirement rather than a courtesy. For a buyer, the limitations section is the part worth reading first.
In practice
Read the out-of-scope and limitations sections before the benchmarks. If a vendor publishes no model card and cannot describe evaluation methodology, treat that as information about the vendor.