Reading a model card before you download

The page that decides whether it is worth the gigabytes

Open models live mostly on Hugging Face, and each one has a model card. Most people scroll to the download button. The five minutes spent reading the card first will save you an evening.

The name, decoded. Qwen3-8B-Instruct is the maker’s own release. TheBloke/Something-GGUF or bartowski/Something-GGUF is somebody else’s conversion of a model into a runnable format: legitimate and useful, and worth knowing it is a conversion. Someone/Model-uncensored or -abliterated is a modified version with safety training removed, which is a decision you should make deliberately rather than by accident.

Base or instruct. Stated on the card. A base model completes text and will not hold a conversation. Almost everybody wants the instruct or chat version.

The licence. A field at the top. apache-2.0 and mit are permissive. llama3.1, gemma, and similar named licences are custom terms with conditions. cc-by-nc-4.0 means non-commercial. This is the next lesson but one, and it is the field most worth reading before you build anything on the model.

Size and file format. .safetensors is the modern format and it is the one to prefer, because it stores plain arrays of numbers. The older .bin weights use Python’s general object-serialisation format, which reconstructs objects by running code as it loads, so a hostile file can run commands on your machine. .gguf is the format for llama.cpp, Ollama and LM Studio: the one you want for running on a laptop.

Context length. Stated in the card and in the config. Note that a converted GGUF may be built with a smaller default, and that your runner may cap it lower again.

Language coverage. Good cards list the languages the model was trained on, with proportions. This is the fastest way to shortlist models for non-English work, and it is far more informative than any benchmark.

The parts that are marketing

Benchmark tables in the card. Self-reported by the maker, run under their chosen settings. Directionally useful within a family (the 32B is better than the 8B) and not comparable across makers.

Download counts and likes. Weak signals dominated by tutorials and by whichever conversion appeared first. A model with 2 million downloads is not necessarily better than one with 40,000; it is better known.

”Comparable to GPT-4”. Every release says this. Some are telling the truth for some tasks. Run your own prompts.

The three things a card usually will not tell you

What it was trained on. Almost no open-weights model publishes its training data. This means you cannot check for contamination, cannot assess bias from the source material, and cannot answer a customer who asks whether their data was in it. AI2’s OLMo family is the notable exception, publishing data and code, and it is worth knowing that the exception exists.

How it behaves on your language in practice. Listed coverage is not measured quality. Test it.

Whether the safety training survived the conversion. Quantised community builds occasionally behave differently at the edges from the maker’s release.

Fine-tunes and merges

A large share of what is on Hugging Face is not an original model but somebody’s adjustment of one: fine-tuned on a particular kind of text, or merged from two or three others. Some are excellent and fill real gaps: a model tuned for a specific language, or for a professional domain, or for a writing style.

They also carry two risks worth knowing about. A fine-tune can lose capability elsewhere while gaining it on its target, and the card rarely measures that. And merges are produced quickly, in volume, often without evaluation of any kind. Prefer the maker’s own release unless a fine-tune solves a problem you actually have, and then test it against the base model on your own prompts rather than on its card.

A five-minute checklist

1. Is it the instruct version?
2. What is the licence, and does it permit what I intend?
3. Is there a GGUF conversion, and what quantisations does it offer?
4. What context length, and does my machine have memory for it?
5. Is my language listed?
6. When was it released? A model from eighteen months ago is usually beaten by something a third of its size now.
7. .safetensors rather than .bin.

Then download the smallest quantisation that fits comfortably, try your twelve prompts, and only fetch a bigger file if the small one nearly works. A 20 GB download you abandon is an evening and, on a metered connection, real money.