Reading a Model Card
How to scan the spec sheet without pretending every benchmark number is holy scripture.
A model card is the closest thing an AI model has to a nutrition label. It tells you what the model claims to be good at, what it was tested on, what license applies, and how much pain you should expect before it runs.
Scan in this order
Task fit. Is the model for chat, code, embeddings, image understanding, speech, reranking, or classification? A great chat model can be the wrong tool for search.
Context window. This is how much text the model can consider at once. Bigger is useful, but it does not magically make the model care about every token equally.
License. This decides whether your employer can use it. "Open weights" and "open source" are not the same thing.
Size and format. 7B, 70B, GGUF, safetensors, Q4, Q8. These terms decide whether the model fits on your laptop, needs a GPU, or belongs in a cloud bill.
Benchmarks. Useful for sorting, dangerous for worship. Benchmark wins can come from narrow test fit, training contamination, or measuring something your app does not need.
Known limits. The boring section is often the most honest section. If the author says the model is bad at long reasoning, function calling, or multilingual tasks, believe them until your evals prove otherwise.
The trap
Do not pick a model from vibes and leaderboard screenshots. Pick a model by matching the card to your actual workload, then run a small eval on examples from your own app.