DigUpModel docs: Google AI — EmbeddingGemma. DigUp project: ARahim3/DigUp. Other apps (e.g. Aether) may also use EmbeddingGemma 2 — this page is DigUp-focused, not a competitor teardown.
What EmbeddingGemma 2 is
EmbeddingGemma is Google’s embedding model family for turning inputs into vectors you can compare. DigUp uses EmbeddingGemma 2 so a typed query can rank photos, PDF pages, document passages, audio windows, and video frames together — not only keyword hits.
- Approximate scale cited in DigUp materials: ~740M parameters; ~865 MB download for DigUp’s quantized on-device package.
- Shared dimensionality commonly described as 768-d for DigUp’s multimodal index.
- Runtime: llama.cpp (GGUF Q8-style weights), Metal on Mac; CPU or GPU (e.g. Vulkan) on Win/Linux after a speed check.
One index, many modalities
| Input | How DigUp uses the model | You get |
|---|---|---|
| Text query | Embed query with text tower at search time | Nearest neighbors + keyword boost |
| Images / screenshots | Visual embeddings (OCR text on Mac for scans/screenshots) | Find by what’s depicted |
| PDF pages | Page-level meaning (+ text / OCR where available) | Open to the right page |
| Audio | ~30s windows in the shared space | Jump near the moment |
| Video | Frame / moment embeddings | Jump near the moment |
Local-first implications
- First launch downloads the model (~865 MB). After that, indexing and search run on device.
- No DigUp account — privacy model is local index + optional update check (can be disabled per upstream docs).
- If network blocks the download (Mac 0.6.1) — Settings → Storage → Use Files… to supply the model files DigUp expects.
- Hardware — Apple Silicon Metal on Mac; Win/Linux pick GPU when faster and equivalent, else CPU (CPU photo indexing is slower).
Sourced numbers (cite date 2026-10-15)
| Number | Context | Where to verify |
|---|---|---|
| ~865 MB | First-run model download size DigUp documents | README / release notes |
| ~740M params | EmbeddingGemma 2 scale (Google / DigUp materials) | Google AI docs |
| 768-d | Shared vector space DigUp describes | Upstream README |
| ~3,716 stars | ARahim3/DigUp popularity signal | GitHub API |