Where are HuggingFaceEmbedding models stored?

[DontEatMyPopcorn]

I cannot identify where in my storage these embedding models are being stored.

For example, when I use the following for embeddings, it installs a new model on my computer.
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
Settings.embed_model = HuggingFaceEmbedding(model_name="Qwen/Qwen3-Embedding-8B")

I cannot find where it is located. There is information about each model in ~/.cache/llama_index/ but the model (i.e., 15 GB of data) is not stored here.

I'm on Fedora Workstation 43.

Any idea where the actual model is stored?

[dosubot[bot]]

Hi @DontEatMyPopcorn! I'm Dosu and I’m helping the llama_index team.

On Fedora (Linux), LlamaIndex's HuggingFaceEmbedding defaults to storing models in /tmp/llama_index, not in the standard HuggingFace cache at ~/.cache/huggingface/hub [1] [2].

That's likely where your ~15 GB model is. Note that /tmp is typically cleared on reboot, so the model may get re-downloaded after a restart.

You can control this with:

• Per-instance: pass cache_folder="/your/preferred/path" to HuggingFaceEmbedding(...) [3]
• Globally: set the LLAMA_INDEX_CACHE_DIR environment variable [4]

For example, to use the standard HuggingFace cache location:

Settings.embed_model = HuggingFaceEmbedding(
    model_name="Qwen/Qwen3-Embedding-8B",
    cache_folder="~/.cache/huggingface/hub"
)

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[DontEatMyPopcorn]

First, I checked my /tmp/ directory. There is no /tmp/llama_index directory.

Next, I rebooted by computer and ran the embedding model. It ran without re-downloading, so it must not have been cleared on reboot. My overall storage space also still appears to be holding the 15 GB.

Where else could the model be stored?

[dosubot[bot]]

Good catch — since the model persisted across a reboot and /tmp/llama_index doesn't exist, it's very likely that SentenceTransformer (which LlamaIndex uses under the hood) ended up storing the model in the standard HuggingFace cache directory: ~/.cache/huggingface/hub/ [1].

This can happen depending on the version of sentence-transformers — some versions may handle the cache_folder parameter differently or fall back to the HuggingFace Hub default.

You can verify by checking:

du -sh ~/.cache/huggingface/hub/models--Qwen--Qwen3-Embedding-8B

You might also check the SENTENCE_TRANSFORMERS_HOME or HF_HOME environment variables if you have them set, as those would override the default location:

echo $HF_HOME
echo $SENTENCE_TRANSFORMERS_HOME

In short, ~/.cache/huggingface/hub/ is the most likely location for your 15 GB model files.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[DontEatMyPopcorn]

I found it. The models were actually in ~/.cache/llama_index
I'm not sure how I made that mistake.

[rehan243]

The Hugging Face embedding models are stored in the Hugging Face cache directory, not directly in ~/.cache/llama_index/. By default, Hugging Face Transformers library stores models in ~/.cache/huggingface/transformers. You can verify this by checking the directory contents after downloading a model.
We use Hugging Face's from_pretrained method under the hood, which follows this caching behavior. You can also specify a custom cache directory using the cache_dir parameter when initializing HuggingFaceEmbedding, like so:

Settings.embed_model = HuggingFaceEmbedding(
    model_name="Qwen/Qwen3-Embedding-8B", 
    cache_dir="/path/to/custom/cache"
)

This can be useful for managing storage or testing different model versions.