Model card

A model card is structured documentation accompanying a released model — what it does, what it was trained on, its evaluation results, intended uses, and known limitations.

Model cards (Mitchell et al., 2018) became a de facto standard by 2026 for both open-weight and closed-API model releases. A good card lists: training data composition, architecture, evaluation scores on standard benchmarks, intended use cases, out-of-scope uses, known biases and failure modes, license, and a contact for issues. Hugging Face requires them; major labs (OpenAI, Anthropic, Google, Meta) publish detailed system cards alongside each release. Reading the model card is the first step before adopting a new model in production — it tells you what the model is for, what it isn't, and how the lab measured it.

Common mistakes

• Comparing benchmark numbers across providers without checking the eval methodology.
• Assuming "strong on benchmark X" means "strong on your task" — model cards rarely match your domain.

FAQ

Related terms

• Evals (LLM evaluations) — Evals are systematic tests that measure how well a language model or LLM-powered system performs on a defined task using a golden set of inputs and reference outputs.
• Fine-tuning — Fine-tuning updates a pretrained model's weights on task-specific data, baking the new behaviour into the model rather than relying on prompts.
• Guardrails — Guardrails are deterministic checks layered around a language model to prevent unsafe, off-topic, or non-compliant outputs from reaching the user.

Sources

• Mitchell et al. 2018 (arXiv)

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/model-card.md.