EmbeddingGemma 2: What It Is and How DigUp Runs It

EmbeddingGemma 2 is the multimodal embedding model released by Google DeepMind in October 2026. Available under the open Apache 2.0 license on Hugging Face, it places text, pictures, audio, and video into one single vector space. DigUp uses it locally to find matching files by meaning across your computer.

Shared 768-dimensional space

Every file segment processed by EmbeddingGemma 2 receives a vector in the same 768-dimensional space. Because all media types share this space, a search query can directly rank a video frame against a PDF page. A photograph of a dog on a beach maps near the text “a dog on the beach”, and an audio recording of a person speaking about sleep maps near “where they talk about sleep”.

Key characteristics of the model include:

• Multimodal coverage: It embeds images, PDFs, text documents, code, audio, and video frames.
• Audio segments: The model listens to audio in 30-second windows.
• Language support: It handles more than 100 languages. Google notes that the model is not equally strong in every language.
• Exact words: Because vector embeddings do not retain exact numbers like invoice codes, DigUp supplements the model by boosting exact-word matches.

How DigUp processes files

DigUp converts your documents into vectors through a local pipeline:

1. Built-in readers convert files from chosen folders into pictures, text, and sound.
2. EmbeddingGemma 2 runs through llama.cpp inside a helper process that exits when indexing completes.
3. Vectors for every picture, page, passage, video frame, and 30-second audio clip are saved into a single SQLite database alongside a keyword index.
4. When you search, the text part of the model (about 250 MB) embeds your search words in memory and compares them with stored vectors and exact terms.

Running the model locally

DigUp downloads ggml-org’s 8-bit build of EmbeddingGemma 2 once during initial setup. The total download is 865 MB, divided into 310 MB for the text component and 555 MB for the image and audio encoders. The index notes which model version created its vectors, and updates to llama.cpp are checked against stored reference vectors.

Execution differs slightly by operating system:

• On Apple Silicon Macs, the model runs on the Mac GPU using Metal.
• On Windows and Linux, DigUp runs on the GPU using Vulkan when setup benchmarks show it is faster than the processor and delivers identical results, falling back to the CPU otherwise.

Try it in DigUp

To see the model running on your local files, download the app from the download page or read how search works on the DigUp home page.

Gemma is a trademark of Google LLC. This guide is independent and is not affiliated with Google.