Model cards and system cards

The document a maker publishes with a model. Read it for what it says, then read it again for the fields it declines to fill in.

Part of the Trust and evals track on lAItest.

The most useful thing in a model card is often the field the maker left blank.

An absence is information.

What the two documents are

A model card is what ships with the weights or the model id: what it is, roughly how it was built, what it was measured on, which licence it carries and where it is known to fail. A system card covers the deployed thing rather than the artefact, including the safety evaluations and red-team findings behind the behaviour that shipped. Neither is an audit. Both are written by the party with an interest in the result. Both are still the best starting point available.

Read them for precision, then for gaps

Cards reward slow reading. One open-weight card cites a maximum generation length used in reasoning evaluations that is a larger, different number from the model’s API output cap — quote the wrong one and your capacity planning is wrong. One lab writes "open weight" consistently and never "open source", which is a deliberate choice rather than a style preference. And plenty of cards simply omit a context window or a price, which tells you the maker has not committed to a number.

Published, and not published, as of Sep 2026

A common misconception

Commonly believed: If a number is not published, it is probably about the same as the competition.

Actually: An unpublished number is a number the maker has not committed to. Aggregator sites fill those holes with guesses, the guesses get copied, and eventually they are quoted back as facts with no source under them. The honest handling is to leave the field empty and say that it is empty, which is what a careful index does and what most comparison sites do not.

In one sentence

Read the card for what it says, then read it again for what it declines to say.