For anyone exploring model downloads
Which model download should I choose?
Understand what you are downloading, which files your software can open, and what to check before you start.
Updated
Start with the software, then choose the file
A model’s “weights” are the numbers it learned during training. Downloading them is not the same as installing an app: you also need software to run them. Different software accepts different file formats.
| If this describes you | Your next step |
|---|---|
| You have not chosen software yet | Start with the model’s information page, called its model card. Look for its instructions on how to run the model before choosing a download. |
| Your software asks for a GGUF file | GGUF is a file format. Follow the GGUF download link from the original model card and check that your software supports this particular model and file version. |
| Your software asks for the original weights | Use the original model download page and its instructions. It may need several files together, including settings and files for converting text into numbers. |
| You see several sizes and do not know which will fit | Read the software’s requirements and the model card. Download size tells you how much data you will receive; it does not tell you all the memory the model needs while running. |
More detail and a worksheet
Ready to try this yourself? These extra checks and a worksheet help you keep track of what you used and what happened.
Check what your software can open
The software that runs a model is sometimes called a runtime. A supported file format is only one requirement: the software must also understand that particular kind of model.
- Find instructions that name your exact model and software version.
- Follow the original model card’s links so you know which download belongs to the release.
- If the instructions are unclear, resolve that question before downloading a large file.
Check memory and usage terms
Some files use fewer bits to store the model’s numbers. This is called quantization; it can reduce file size, but you still need to check memory use and results for that exact version.
- Allow for the model, the text it is processing and the software itself. Measure your actual setup during a small trial.
- Write down the file variant you chose. A score reported for a different variant may not describe yours.
- Read the licence: it sets the terms for using or sharing the files. Different models can have different licences.
Save enough detail to repeat your trial
A download page can change. Recording a revision — an identifier for one exact version — helps you find the same files again.
- Save the download link, revision and filenames. If a checksum is provided, save it too; it is a fingerprint used to check a file’s contents.
- Note your software version, computer and settings, including how much text the model can handle at once (its context limit).
- Try one small task and save the question, answer and time taken before planning a larger test.
Keep a download note
Copy this worksheet. If you cannot find a detail, write “not provided” rather than guessing.
Model name and size: Model information page (model card): Download page: Exact version identifier (revision): Downloaded filenames: File fingerprints, if provided (checksums): File format, such as GGUF: Smaller-number variant, if used (quantization): Licence link: Software name and version: Instructions confirming this model is supported: Computer and available memory: Settings, including maximum text length (context): Trial question and expected answer: Actual answer and time taken: Where I saved the files and result:
Sources and next steps
A model card is the publisher’s information page for a model. Follow it for the exact download instructions, licence and known limits. External technical sources are in English.