Vibeleaderboard
← All Intel
Intel / article

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Source
deepmind.google
Date
From the Daily Brief

Google DeepMind released EmbeddingGemma 2, an open embedding model with 740M parameters built on the Gemma 4 architecture. Its main change is scope: text, code, images, audio and video all land in a single embedding space, so one index can answer a text query with a video clip or a screenshot. The model is small enough for on-device use, which Google frames around private search and retrieval-augmented generation without sending documents to a hosted embedding API. The Apache 2.0 license permits commercial use. Teams that currently run separate text and image embedders, or pay per call for hosted multimodal embeddings, have a local option to benchmark against their own retrieval set. Google's announcement does not substitute for that test, since retrieval quality varies widely by domain.

Read the 2026-10-07 Brief →

Context

models turn content into lists of numbers so that similar items sit close together, which lets an app search by meaning instead of exact words. Google DeepMind says EmbeddingGemma 2 extends that idea from text to code, images, audio and video in one shared space, so a text query can search audio recordings or locate moments in video. The model has 740 million parameters, builds on the Gemma 4 architecture, uses the Apache 2.0 license, and is built for on-device use.

Google reports a code-search score of 78.68 on MTEB Code, up 9.92 points from the first EmbeddingGemma, and describes it as leading among multimodal embedding models under one billion parameters. These are the developer's own benchmark claims, with full results in a model card. With , Google says the model needs about 191 MB of active memory for text only and about 567 MB for the full multimodal version on a Pixel 11 Pro.

Key takeaways · AI-distilled
  • The model is modular: Google says text-only workloads need as little as 270M parameters, with an optional 170M vision encoder and 300M audio encoder added only when you need full multimodal support.
  • Matryoshka Representation Learning lets developers truncate output vectors from 768 dimensions to 512, 256 or 128, which Google says cuts local vector-database storage and memory by up to 6x.
  • On a Pixel 11 Pro with quantization, Google reports roughly 191MB of active RAM for text-only weights and about 567MB for the full multimodal model.
  • The is 8K tokens, four times EmbeddingGemma 1, enough per Google for up to 5.5 minutes of audio, 29 images or 58 video frames, or interleaved mixes of them, in one input.
  • Google reports code retrieval improved most: MTEB Code rose from 68.76 to 78.68 over EmbeddingGemma 1, while multilingual text performance held steady, pitching it for local codebase indexing and coding- retrieval.
Terms in this piece · Glossary
  • embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
  • quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Read the source deepmind.google
Recommended reads
Comments

Checking sign-in…

Loading comments…