Google EmbeddingGemma 2: How One Model Embeds Text, Images, and Audio
Google's EmbeddingGemma 2 maps text, images, video, and audio into one 768-dim vector space under Apache 2.0. Here's how it actually works.
What is EmbeddingGemma 2?
EmbeddingGemma 2 is an open embedding model from Google DeepMind that converts text, code, images, video, and audio into vectors living in the same 768-dimensional space. Released under Apache 2.0 with weights on Hugging Face, it lets you compare a voice memo to a video clip or a text query to a photo using one model instead of stitching together separate encoders. It totals 740 million parameters but loads modularly, so you only pull in the encoders for the modalities you actually use.
TL;DR
- EmbeddingGemma 2 puts text, code, images, video, and audio into one shared 768-dimension vector space, so any modality can be compared against any other with a single model.
- The model is modular by design: the text backbone is 270 million parameters, with a 170 million parameter vision encoder and a 300 million parameter audio encoder loaded only when needed, keeping total size at 740 million parameters.
- Matryoshka truncation lets you cut the 768-number vector down to 512, 256, or 128 dimensions, shrinking storage up to six times while keeping the most important information near the front of the list.
- In hands-on testing, VRAM use for the text-only modality stayed under 2GB, making it practical to run on modest hardware or even CPU.
- A real test against six AI-generated images showed the model correctly matched every text query to its intended image, including abstract phrasing like “something delicious” correctly landing on a curry photo.
- Cutting the same image-search vectors from 768 down to 256 dimensions (a 3x storage reduction) did not change a single top result, showing Matryoshka truncation holds up in practice at that cut point.
- On Google’s own code search benchmark, EmbeddingGemma 2 scored around 78, ahead of models several times its size, like the 1.5 billion parameter INF-retriever.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
How does EmbeddingGemma 2 turn different media into the same kind of vector?
The core trick is architectural. Text runs through a tokenizer, images and video frames run through a vision encoder, and audio runs through an audio encoder. All three paths converge into one shared text backbone, so no matter what goes in at the top, what comes out at the bottom is the same kind of object: a list of 768 numbers.
That shared output format is what makes cross-modal search possible. A text question like “a wild animal at sunset” and a photo of an actual animal at sunset end up as vectors sitting close together in that 768-dimensional space, because both were mapped into the same coordinate system rather than two separate, incompatible ones. Similarity between any two items, text-to-text, text-to-image, or audio-to-video, becomes a single, consistent calculation: how close their vectors sit to each other.
This is different from bolting a vision model and an audio model onto a text model and hoping their outputs line up. EmbeddingGemma 2 is trained so all modalities land in the same space on purpose.
What does “modular” actually mean for VRAM and deployment?
Despite being described as one model, EmbeddingGemma 2 is not a single monolithic blob you have to load in full every time. The text-only component is 270 million parameters. The vision encoder adds roughly 170 million, and the audio encoder adds around 300 million. Added together, the full multimodal model comes to 740 million parameters, but you only load the pieces you need.
If a project only needs text embeddings for search or retrieval, you load the 270 million parameter text path and skip the rest. In hands-on testing using the sentence-transformers library on top of transformers, running text-only embedding generation consumed just under 2GB of VRAM, light enough to run comfortably on a single consumer GPU or even a CPU. That modularity matters for anyone deploying on constrained hardware: a mobile app doing on-device semantic search doesn’t need to carry an audio encoder it will never call.
What is Matryoshka truncation and why does it matter?
Matryoshka representation learning, named after Russian nesting dolls, trains the model so the most meaningful information is packed toward the front of the embedding vector. Because of that ordering, you can truncate a vector down to its first 512, 256, or 128 numbers and still retain most of its usefulness, instead of needing the full 768 numbers every time.
This matters directly for cost. A vector database storing millions of embeddings at 768 dimensions versus 128 dimensions is a roughly six-fold difference in storage and memory footprint. According to Google’s published guidance, quality holds up well down to 256 dimensions, but drops off at 128, particularly for images, video, and audio, where more dimensions are needed to preserve the extra information those modalities encode.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
A direct test bore this out: six images were embedded alongside six text queries (“a cat on a mountain,” “something delicious,” “a person sleeping in a car,” and others), with similarity computed first at the full 768 dimensions and then again after truncating everything to 256. Every single query matched the same top image in both cases, despite using one-third the storage. That’s a useful data point for anyone trying to decide where to make the quality-versus-storage tradeoff.
How accurate is EmbeddingGemma 2 in practice?
Textual similarity tests showed the model ranking sentence pairs in exactly the order expected: a pair expressing the same idea in different words scored highest (around 0.956), a loosely related pair scored in the middle (around 0.863), and an unrelated pair scored lowest (around 0.733). Notably, even the “unrelated” pair didn’t score near zero. Google’s own documentation describes this as expected behavior for the model, meaning the relative ranking between scores is what matters, not the absolute number.
For cross-modal search, six text queries were matched against six images with no shared keywords between query text and filenames. Phrases like “a wild animal at sunset” correctly retrieved an image of a hyena, “two people staring at each other” retrieved a photo described as a gaze, and “something delicious” retrieved a curry dish despite the word “curry” never appearing in the query. Scores for text-to-image matches were lower than text-to-text matches, generally between 0.65 and 0.80, which is typical when comparing across modalities. The important signal is the gap between the top result and the runner-up, not the raw number.
On a published benchmark measuring code search quality against model size, EmbeddingGemma 2 scored near 78, placing it ahead of considerably larger models, including a 1.5 billion parameter retriever, a strong result for a model with a 270 million parameter text backbone.
Is EmbeddingGemma 2 worth using over a text-only embedding model?
If a project only ever needs to embed and search text, a smaller dedicated text embedding model may be simpler to deploy and just as effective. But EmbeddingGemma 2’s real value shows up the moment a project needs to search across modalities: finding a video by describing it in words, matching a text query to an audio clip, or searching a photo library with natural language. Having one shared vector space instead of multiple disconnected embedding models removes a significant amount of integration complexity.
The combination of an Apache 2.0 license, modular loading, low VRAM requirements for text-only use, and Matryoshka truncation for storage efficiency makes it a practical option for teams building retrieval systems without heavy infrastructure. It’s less compelling for teams that need only one modality and have no plans to expand beyond it.
Frequently Asked Questions
What license is EmbeddingGemma 2 released under?
It’s released under Apache 2.0, and the model weights are publicly available on Hugging Face under Google’s account.
How many dimensions does EmbeddingGemma 2 use, and can that be reduced?
The native output is a 768-dimension vector. Using Matryoshka truncation, that can be reduced to 512, 256, or 128 dimensions, with quality staying strong down to 256 and degrading more noticeably at 128, especially for non-text modalities.
Do I need to load the vision and audio encoders if I only use text?
No. The model is modular. The text-only component is 270 million parameters, and the vision (roughly 170 million) and audio (roughly 300 million) encoders are loaded separately only when those modalities are needed.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How much VRAM does EmbeddingGemma 2 require?
For text-only embedding generation, hands-on testing showed VRAM usage staying under 2GB, making it light enough for most consumer GPUs or CPU-based inference.
What context length does EmbeddingGemma 2 support?
It supports an 8K token context window for text input.
