How to Run EmbeddingGemma 2 On-Device

For EmbeddingGemma 2. Difficulty: intermediate. Time: PT30M. Commands below are based on official docs, not yet tested.

Prerequisites

A machine that can load a ~0.74B embedding model (exact RAM/VRAM not in our seed extract)
For the Mac app path, a Mac (DigUp is a Swift on-device search app)

Decide between the model card and DigUp. For raw embeddings, start at https://huggingface.co/google/embeddinggemma-2. For Mac file search that already wraps the model, DigUp is a verified app (https://github.com/ARahim3/DigUp, MIT, Swift, 1,712 stars, created 2026-10-08).
Read the Hugging Face model card. Copy transformers / sentence-transformers / other runtime snippets from the card. This page does not invent a code sample.
Check the Apache-2.0 LICENSE on the card. Confirm the license still matches what we recorded on 2026-10-13. This is not legal advice.
If using DigUp, install from its GitHub README. We have not run DigUp. Follow ARahim3/DigUp. Do not assume App Store availability; that was not verified.

Based on official docs, not yet tested.

Out-of-memory errors. We do not have a sourced min-RAM figure. Reduce batch size only using guidance from the official card, or use DigUp’s own docs.
Confusing EmbeddingGemma 2 with BGE-M3 snippets. They are different models. See the comparison page before mixing checkpoints.

Did you test this on a specific device?

No. lastTested is empty on purpose.

Is DigUp the same as running the model in Python?

No. DigUp is a Mac semantic file search app built on the model. The Python/transformers path is whatever the official card currently shows.

Can I use this commercially?

The model card lists Apache-2.0; DigUp lists MIT. Read both licenses.

Last verified 2026-10-13 · Last updated 2026-10-13· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.