Model hubs and formats

Model hubs are where open models are published, and file formats decide whether a downloaded model will load in your tool. Every local AI setup touches both, usually without noticing, so it is worth knowing what you are looking at before you download several gigabytes.

A hub hosts model repositories much like a code host hosts projects. Each model page has its files, a model card describing how it was trained and what it is meant for, and a license. Popular models are re-uploaded by other people in quantized versions, which are smaller files meant for running on ordinary hardware. Those re-uploads are convenient, but they are only as trustworthy as the person who made them.

File formats come next. GGUF bundles a model, its tokenizer and settings into one file and supports many quantization levels; it is what llama.cpp-based apps expect. Safetensors stores weights in a way that cannot run code when loaded, which makes it the safer choice for full-size models and image models. Older pickle-based files can execute code on load and should be treated with caution.

Read the model card before you download. It tells you the intended use, known limits, the languages covered and how the model was evaluated by its makers. It also links to the license. Some models are open for any use, some restrict commercial use or require attribution, and some ask you to accept terms or share contact details before downloading.

Featured local AI tools: model hubs and formats

SponsoredModel hubs and formats This position is open Put your tool at the top of the model hubs and formats page. Paid spots are always labeled Sponsored. From $99 a month. Get featured

What a featured listing costs ยท from $99 a month

5 local AI tools in model hubs and formats

GGUF

by ggml-org

The single-file model format used by llama.cpp and many local apps, with built-in support for quantized weights.

  • Open source

safetensors

by Hugging Face

A file format for storing model weights that loads fast and cannot run hidden code when opened.

  • Open source

Hugging Face Hub

by Hugging Face

The largest public host for open model weights, datasets and demos, where most local models are first published.

ModelScope

by Alibaba

A model hub run by Alibaba that hosts open models and datasets, including many released by Chinese labs.

Kaggle Models

by Google

Google's model repository on Kaggle, with open models ready to download or try in notebooks.

Choose the file size with your hardware in mind. Hub pages for quantized models usually list several versions of the same model. Smaller files load on more machines and lose some quality; larger ones are closer to the original. Starting in the middle and adjusting is a reasonable habit.

Model versions change more often than people expect. A publisher may update weights, fix a tokenizer or change the license on an existing page, and quantized re-uploads may lag behind. When results suddenly differ from what you saw before, check whether the file you are using matches the version described on the current model card, and pin a specific version for any work you need to repeat.

Finally, keep track of what you downloaded and from where. Model files are large and easy to forget about, and knowing the exact source and version makes it possible to check licenses later or reproduce results.

Questions about model hubs and formats

What does quantized mean on a model page?

The weights have been stored with fewer bits, so the file is smaller and needs less memory. Quality drops a little, and more aggressive quantization drops it more. Most people running models locally use quantized files.

Is it safe to download models from re-uploaders?

Often, but not always. Prefer the original publisher or a well-known uploader, prefer GGUF or safetensors files over older pickle formats, and check that the license matches the original model.

Which format should I download?

It depends on your tool. Apps built on llama.cpp, including many desktop apps, use GGUF. Python libraries and image interfaces usually load safetensors. Check your tool's documentation before downloading large files.

Why do some models ask me to accept terms first?

Their publishers require agreement to a license or use policy before download. Accepting creates obligations about how you use the model, so read the terms instead of clicking through.

Related categories