embeddinggemma-2

embeddinggemma-2

Google DeepMind's open embedding model that places text, code, images, video and audio in one 768 number vector space, built to run on laptops and phones. Suits developers building search and retrieval.

Made by google on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: EmbeddingGemma 2 is an open embedding model from Google DeepMind. It turns text, code, images, video and audio into vectors in one shared 768 number space, so different kinds of media can be compared with each other. It joins a 270M text model with a 170M vision encoder and a 300M audio encoder, for about 740M parameters in all.

Who it suits: Developers building search, retrieval for chat models, classification or clustering, especially where the work must run on a laptop or phone.

Industry | Who uses it | How they use it | Example
Software | Search engineers | Index text, code, images, video and audio in one vector space | Match a text query to a video clip
Mobile apps | App developers | Run semantic search on the device | Search a user's notes and photos together
AI products | Retrieval builders | Fetch passages for a chat model | Embed documents and questions with matching prefixes
Data work | Analysts | Group or label items by meaning | Cluster a pile of support messages

Skip it if you need a model that writes text, because it only produces vectors.

Where to use it:

• Python programs that use the transformers library, which the page lists
• Laptops and phones, which the card names as targets
• Vector search and retrieval pipelines

Use case | Who | What they get back
Semantic search across media | Developers | Ranked matches for text, image, video and audio items
Classification and clustering | Data teams | Items grouped by meaning
Smaller vector storage | Teams with large indexes | Shortened vectors, up to 6x less storage per the card

How to set it up:

• Python
• The Hugging Face tools, to fetch the files
• The transformers library in your own code
• About 1.5 GB of disk for BF16 weights, or about 372 MB at 4 bit as an estimate

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Load the model in your own code with the transformers library.

pip install -U huggingface_hub
hf download google/embeddinggemma-2

How to use it: Pass text, images, video or audio through the model to get 768 numbers per item. Add a short instruction prefix that fits the task, then compare vectors to find matches.

Tool | What it does
hf download | Copies the repository files to your computer
Task prefix | A short text instruction that steers vectors for search, classification, clustering or similarity
Truncated vectors | Cuts vectors to 128, 256 or 512 numbers, then re-normalises them
Loadable encoders | Lets you load only the vision or audio encoder you need

Example: Embed the words "red running shoes" and a folder of product photos, then rank the photos by how close their vectors are to the text. The closest photos should be the ones that show red running shoes.

For newcomers and experts: If you are new to this, think of an embedding as a list of numbers that places similar items near each other. Embed two sentences and one unrelated sentence, then compare how close each pair is.

If you know the field, the text model has 24 layers, a 512 wide model dimension, a sliding window of 1024 tokens and a 5 to 1 local to global attention mix. Mean pooling feeds a 512 to 768 projection. Matryoshka training allows truncation to 128, 256 or 512 dimensions.

Text, image, video or audio goes in
The matching encoder reads it
A 768 number vector comes out
You compare vectors to find matches

Title: MTEB mean task scores, version 2 against version 1
Code, version 2: 78.68 points
Code, version 1: 68.76 points
Multilingual, version 2: 61.36 points
Multilingual, version 1: 61.15 points

Good to know:

• The context window is 8,192 tokens.
• The reported scores come from the full precision checkpoint, so quantized copies may score differently.
• Native vectors have 768 numbers, with truncation to 128, 256 and 512 supported.
• License: apache-2.0.
• Cost to run: none named, only your own hardware.
• Data it reaches: the text, images, video and audio you embed. The source names no account or outside service.
• Not for: the source names none.

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: transformers
• Parameters: 744M
• Weights: about 1.5 GB in BF16; about 372 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 503,609
• Likes: 3,347
• License: apache-2.0
• Quantized versions: 139
• Fine-tunes: 170
• Tags: transformers, safetensors, embedding_gemma2, feature-extraction, embedding, sentence-transformers, multimodal-embedding, multimodal, vision, audio, video, image-feature-extraction

Links

• Model: https://huggingface.co/google/embeddinggemma-2
• Files: https://huggingface.co/google/embeddinggemma-2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download google/embeddinggemma-2

Then follow the usage in its model card. Nothing in this kit runs by itself.