What is a model card?
The spec sheet that travels with a model. Before you download anything, the card is where the honest facts about it live.
A model card is the document that describes a specific model: its size, what it was trained on, what it is meant for, how it performs, where it fails and under what license it ships. It is the closest thing a model has to a spec sheet. Reading it first tells you whether the model fits your hardware and your job, and reading it after a surprise tells you whether the surprise was documented.
The facts before the download.
Models travel as opaque files, and a downloadable model gives you no interface to quiz about its own properties. The card is the publisher's answer in advance: how many parameters, trained on what kinds of text, tuned for which tasks, tested on which benchmarks, known to fail where. One of the oldest numbers on it is perplexity, and what perplexity is covers what that score measures and what it leaves out. Some of it is numbers, like the parameter count that predicts the memory bill; some of it is plain description of intent and limits that no benchmark captures.
What to read first.
Three parts repay the attention. Size and memory cost decide whether the model runs on your machine at all, a question the RAM arithmetic answers once you know the numbers. The intended-use section says what the training aimed at, which matters more than any headline score: a model built for conversation can be mediocre at summarizing documents, and the card usually says so if you read it. The license governs what you may do, and it varies widely across releases that look identical from the outside.
A description, not a guarantee.
A card is written by the model's own publisher, so treat it as a description with an interest, not an independent audit. Most are candid, and the limits sections in particular tend to be, but community testing after a release sometimes finds weaknesses the card never mentions. The practical habit is to read the card, glance at the discussion around the release, and then judge the model on your own documents. That is also what a tool with a fixed roster does on your behalf: whoever picked those models read the cards and checked the claims against the work. The place those cards live is Hugging Face, the model library most downloads start from, and the card is the first thing its pages ask you to read.