What is a GGUF file?
The file format local models ship in. One download, everything the runtime needs, and the reason "try this model locally" is a copy operation instead of a setup project.
A GGUF file is the packaged form of a local language model: its weights, usually already quantized, plus the metadata a runtime needs to load it, all in a single file; that metadata often records what the model card described, so the file carries its own spec sheet. The format grew out of the llama.cpp project and became the standard because it is simple and self-contained. Download one, point a runtime at it, and the model runs on your CPU or GPU with nothing else to configure.
One file, everything included.
Before GGUF, running an open model meant collecting the weights, the tokenizer files, and the configuration separately, and hoping the versions matched. GGUF folds all of it into one binary: the weights, the architecture description, the tokenizer, and the prompt template the model was trained with. The self-containment is what makes local distribution workable, since a model becomes a file you can copy between machines, back up like any document, and load with whatever local runtime you use. The files are possible because the model creators publish their weights; what open-weight means covers that side of the exchange.
Quantization travels inside it.
The format's other convenience is that quantization is baked in at packaging time. A model is usually published as several GGUF files, one per bit depth, and the name carries the level, so a Q4 file of a 7B model weighs about 5 gigabytes while its Q8 sibling takes twice that; the full code, like Q4_K_M, encodes more than just the bit depth. You choose the file that fits your memory budget rather than configuring anything, which is why the RAM question is really a file-choice question.
Why it became the standard.
Adoption followed simplicity. The llama.cpp ecosystem reads it natively, so every tool built on that stack, from command-line runners to desktop research apps, takes the same file. Model hosts list GGUF quantizations alongside the original weights, which gives the format a distribution network no rival managed. The practical result for a researcher is unglamorous and genuinely useful: the barrier between deciding to try a model and running it locally is one download. Those model hosts are their own topic: what Hugging Face is covers the library where these files live and what to read before pulling one down.