EmbeddingGemma 2: An open, lightweight multimodal embedding model

TL;DR Highlight

One open 740M model embeds text, images, audio and video into a single shared vector space

Who Should Read

Backend or app developers building on-device or self-hosted RAG who need to search images, audio and video as well as text

Core Mechanics

Text, images, audio and video share one space
- EmbeddingGemma 2 maps code, images, video and audio into a shared embedding space.
- It is built on the Gemma 4 architecture and released under Apache 2.0.
- It has 740M parameters in total.
Load only the encoders you need
- Text-only workloads need as little as 270M parameters.
- Vision (170M) and audio (300M) encoders are optional add-ons.
MRL cuts vector storage up to 6x
- Matryoshka Representation Learning (MRL, training so the first part of a vector still works when truncated) lets you cut 768 dims to 512, 256 or 128.
- Local vector DB storage and memory drop by up to 6x.
Code retrieval score up 9.92 points
- On MTEB Code it went from 68.76 to 78.68 versus EmbeddingGemma 1.
- Multilingual text performance is said to match EmbeddingGemma.

Community Reactions

Apache 2.0 matters for embedding models, one commenter argued
- Proprietary hosted models can be retired, forcing you to re-embed millions of stored vectors.
- The commenter prefers paying a host while keeping open weights as a fallback.
Text-only users may see no gain, per commenters
- One commenter noted the text benchmarks are identical to the first EmbeddingGemma.
- Another tested with Claude and found it neither better nor faster for text retrieval or clustering.
- They said MuQ-MuLan still seems better for music audio.
MRL doesn't shrink the model weights, one commenter pointed out
- Unlike earlier on-device MatFormers models, only the dimensions shrink, not the weights.
Commenters wanted more comparisons
- Some asked for comparisons with Voyage AI for text and SigLIP2 for images.
- One said they found Voyage models better than Qwen ones.

How to Apply

Compare the text-only path on your existing corpus
- Today, load the 270M text path with sentence-transformers and run the same 20 queries you use now.
- If your service is text-only, check first whether there is any gain.
Shrink vectors with MRL dimensions
- Index 256 or 128 dims instead of 768 in your vector DB and measure the quality change.
- Rebuild the index at the smallest dimension that stays acceptable.
Try it on codebase search
- Plug it into your coding agent's code retrieval and compare hit rate against your current model.
Add encoders only where you need them
- Add the audio (300M) or vision (170M) encoder if you need voice memo or video search.
- Otherwise keep text-only to save memory.

Try It

copy & run

# Run in terminal (the source only says sentence-transformers is supported; check and replace the model ID below)

pip install -U sentence-transformers

# Save as a Python file and run
from sentence_transformers import SentenceTransformer

MODEL_ID = "{{EmbeddingGemma 2 model ID on Hugging Face}}"
model = SentenceTransformer(MODEL_ID, truncate_dim=256)  # MRL: 768 -> 256 dims

docs = ["{{doc1}}", "{{doc2}}", "{{doc3}}"]
query = "{{question}}"

d = model.encode(docs, normalize_embeddings=True)
q = model.encode([query], normalize_embeddings=True)
print(q @ d.T)

# Next: change truncate_dim to 512, 256, 128 and see how the ranking shifts

Terminology

EmbeddingTurning text or images into number vectors so similar meanings sit close together. Think of it as giving each item a coordinate for meaning.

MRL (Matryoshka Representation Learning)Training like a nesting doll so the first part of a vector is still useful on its own. You can store 128 dims instead of 768.

Multimodal embeddingPutting text, images, audio and video into one vector space. It lets you find a video moment with a text query.

QuantizationStoring model weights at lower precision to save memory. Similar to compressing a photo.

On-deviceRunning the model directly on a phone or laptop instead of a server.