Model cards: How to read the label that comes with an AI model
A model card says what an AI model is for, how it was tested and where it fails. Here is what is in one, where the idea came from and why self-reported claims need a second look.

A model card is a short document that ships with an AI model and says what the model is for, how it was tested and where it is likely to fail. Think of it as the label on a medicine box, written for the people who will use the model or decide whether to trust it.
The idea is almost a decade old, but it keeps coming up in news about new releases, including this week’s Gemini 4 Carbon.
A label you can read before you rely on the model
Suppose a company wants a model to sort customer emails. Two models both claim to be “the best”. Without documentation, the team has to run its own tests to find out which one fits.
A model card shortens that work. It states the intended use, the data and tests behind the headline numbers, and the cases the maker already knows are weak. A reader can then see whether the model matches the job before writing any code.
The card does not make a model safe or good. It makes the maker’s claims visible, so others can check them or challenge them.
What a card usually contains

The format has a few standard parts. Hugging Face, the largest public host for open models, says a card should describe:
- the model itself;
- its intended uses and potential limitations, including biases and ethical considerations;
- the training parameters and experimental information;
- which datasets were used for training;
- the evaluation results.
On the Hub, a card is simply the README.md file in a model repository. It is a Markdown text file with a block of structured metadata at the top, written in a format called YAML. Fields such as license, datasets, base_model and pipeline_tag let people filter and compare models. Evaluation scores can also go into the metadata, so the Hub can show them in a results widget.
The base_model field is a good example of why structure matters. It records whether a model is a fine-tune, an adapter, a quantized copy or a merge of other models, so a reader can trace where it came from.
Where the idea came from
The concept was published in the paper “Model Cards for Model Reporting” by Margaret Mitchell and eight co-authors, including Timnit Gebru and Inioluwa Deborah Raji. It appeared on arXiv in October 2018 and was presented at the FAT* conference in January 2019.
The paper’s main argument was about fairness. It proposed that released models should report performance separately for different demographic, cultural and intersectional groups, not as one average. A face detector that works well overall can still work poorly for some groups, and a single score hides that.
The authors showed the format with two example cards, one for a smiling-face detector and one for a toxic-comment classifier. They presented cards as a companion to “datasheets for datasets”, which document the training data.
Hugging Face later built the format into its platform. Its documentation credits the 2018 paper by Mitchell and colleagues as the source for the intended-use and limitations sections.
Where model cards show up in this week’s news
Our report on this week’s new models leans on cards in two ways. For Gemini 4 Carbon, the context window, the long-horizon design and the benchmark scores come from Google’s model card. According to that card, Carbon uses more thinking tokens to test hypotheses and run verification loops, and it reports 65.2% on Terminal-Bench 4.0.
The second way is on the open side. Thinking Machines, MiniMax, Zyphra and Cactus Compute all published weights on Hugging Face, where model pages are built around a card in the format described above.
Closed-model makers also publish launch documents. Anthropic’s post on Claude Fable 5.2 reports scores on FrontierCode 2.0 and OSWorld 2.1, and our report notes these are the company’s own results and were not reviewed by outside organizations. That caveat applies to maker documentation in general.
What a card cannot tell you
A card is a claim by the maker. Nobody has to verify it, and no authority checks that it is complete. Three limits matter in practice.
- Self-reported numbers. Scores are usually the maker’s own, measured with the maker’s setup. Different prompts, thinking settings or tool access can change a result, and the card may not say which were used.
- Missing sections. Many community cards leave out training data, limitations or bias analysis. An empty section is information, but it is not a test result.
- Stale text. A card describes a model at one moment. If the maker changes the weights, safeguards or pricing, the document can fall behind.
Cards also do not replace independent testing. Outside evaluators and benchmark indexes exist because maker reports alone are not enough for high-stakes decisions.
How to read one in two minutes
A quick routine helps when a new model arrives:
- Check the intended use. If it does not match your task, the benchmark scores matter less.
- Check the
licenseand thebase_modelfield, so you know the legal terms and what the model was built from. - Look at how the scores were measured: which benchmark version, which settings, and who ran the test.
- Read the limitations section. Specific limits are a better sign than vague ones.
- Look for a second source, such as an outside evaluation, before relying on the headline numbers.
What to watch next: whether cards for large closed models become more standardized and more detailed, and whether outside reviewers test the claims in them. For now, the quality of a card still varies widely from one release to the next.
Sources
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Daniel ReedAI Research, Safety & Ethics AnalystDaniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…



