A model card typically includes: model details (architecture, version, date), intended use (what the model is designed for and what it shouldn't be used for), training data (description of the training dataset, including any known biases), performance metrics (broken down by relevant subgroups), limitations (known failure modes, edge cases), and ethical considerations (potential harms, mitigation strategies).
Hugging Face popularized model cards by requiring them for all models on their Hub. Quality varies widely — some are detailed technical documents, others are perfunctory placeholders. The best model cards include per-group performance breakdowns (does the model work equally well for different languages, demographics, or domains?), concrete examples of failure cases, and honest assessments of limitations rather than marketing language.
The concept extends beyond models: data cards document datasets (collection methodology, annotation process, known biases), and system cards document entire AI systems (model + post-processing + guardrails + deployment context). Anthropic publishes system cards for Claude releases. These broader documents capture information that model cards alone miss — a model might be safe in isolation but dangerous when deployed with certain tool-use capabilities or without content filters.