The two documents
Modern LLM releases ship two complementary documents. Confusing them is common; reading both is the right discipline.
Model card
Describes the technical model: architecture, training data composition, training compute, hyperparameters, evaluation results, intended use, and license. Hugging Face hosts these as README.md alongside the weights. The most useful sections are usually:
- Architecture — attention type, FFN type, MoE config, position encoding, vocab.
- Training — dataset composition (often vague), training compute, optimization.
- Evaluation — benchmarks the lab wants to highlight (apply benchmark-skepticism stack).
- Limitations — what the lab admits the model is bad at.
- License — what you can do with the weights.
System card
Describes the safety/policy/evaluation story: red-teaming results, refusal behavior, known harms, deployment-specific safety measures. Anthropic and OpenAI publish detailed system cards alongside major releases. These are denser than model cards and often contain the most surprising information.
What to read first
- Model card architecture section — locates the model on axis 1.
- Training section — locates it on axis 2 to whatever extent the lab discloses.
- Evaluation, but skeptically.
- License — before deploying.
- System card — before deploying to users.
What is NOT in the cards is also signal
Cards that omit dataset composition, training compute, or red-teaming results are telling you the lab does not want to disclose those. Cards that omit known limitations are telling you the lab is doing marketing in technical clothing. Read what is missing as carefully as what is present.