How much storage does a local AI server need for model files?

Updated November 3, 2026.

There is no single storage number for a local AI server, because it depends on which models you keep and in how many versions. The reliable method is to add up the size of the files you plan to keep, add the same again for a backup, and leave room for new downloads. This page shows where the files live and which commands report real sizes, so you measure instead of guessing.

Where model files live

Ollama's FAQ gives the default model folder: ~/.ollama/models on macOS, /usr/share/ollama/.ollama/models on Linux and C:\Users\%username%\.ollama\models on Windows. If your system disk is small, set the OLLAMA_MODELS environment variable to a folder on a larger disk. On Linux with the standard installer the ollama user needs read and write access to that folder, and the FAQ gives the chown command for it. Our page on moving Ollama models to a new computer covers copying the folder.

If you download model files from Hugging Face directly, they go to a different place. The huggingface_hub cache documentation says the default cache is ~/.cache/huggingface/hub and that you can change it with HF_HOME or HF_HUB_CACHE. Chat history is stored separately; in the Docker example from Open WebUI's Tailscale guide, its data sits in a volume mounted at /app/backend/data.

Measure what you have

Ollama's command line reference lists ollama ls to show the models you have and ollama rm to remove one. For Hugging Face downloads, the cache guide shows hf cache ls, which prints each cached repository with its size and when it was last accessed. Its example output ends with a total of the space used on disk, and filters such as size greater than 1GB or accessed more than 30 days ago help you find files to remove. The same guide documents hf cache rm and hf cache prune for deleting entries.

For a planning figure before you download, use the file size shown on the model download page. Different quantizations of the same model are different files with different sizes, as our page on GGUF quantization explains, so the quantization you pick is the biggest lever on how much disk you need. Multiply by the number of models and versions you intend to keep.

Keeping the folder from growing

Models you stop using still take space until you remove them. Open WebUI's Ollama guide says that deleting a model from its manager removes it from the Ollama server itself and frees its disk space, and that it disappears for every user of that connection, so check with the other users first. If you only want to hide a model, the same guide describes listing just the model IDs you want to offer instead. The guide also describes an Update All Models button that pulls every model on the server again; it says Ollama downloads only what changed. Check ollama ls afterwards to confirm the sizes.

Plan for backups and free space

A backup of the model folder needs as much space as the folder itself, so budget twice the total, and test a restore as described in our page on backing up and restoring local AI model files. Keep free space for at least the largest model you might pull next. Finally, remember that model files are only part of the picture: exported chats are a JSON file you create yourself, and any document index you build for private search adds its own storage.

Frequently asked questions

Where does Ollama store models?

By default in ~/.ollama/models on macOS, /usr/share/ollama/.ollama/models on Linux, and C:\Users\%username%\.ollama\models on Windows. Set OLLAMA_MODELS to use a different folder.

How do I free space used by a model I no longer need?

With Ollama, run ollama rm followed by the model name. Open WebUI's model manager can also delete a model from the connected Ollama server, which frees its disk space.

Where does Hugging Face store downloaded models?

By default in ~/.cache/huggingface/hub. The HF_HOME or HF_HUB_CACHE environment variable changes it, and the hf cache ls command shows how much space each cached repository uses.

Sources

Sources checked November 3, 2026. Figures that depend on your exact model, runtime or machine are estimates; check the source or test on your own hardware.

Adults only (18+). Private uncensored AI · Terms · Privacy · All guides