# Google presents open 744M-parameter EmbeddingGemma 2 for text, image, audio and video

> The open model maps text, images, audio, video and mixed inputs into one vector space and is described as a practical on-device option.

- Status: official (From the company, authors or authority themselves, with a link to their own page.)
- Topic: Models
- Published: 2026-11-01T17:34:56+00:00
- Updated: 2026-11-01T17:34:56+00:00
- Web page: https://ai-news.ikangai.com/story/google-presents-open-744m-parameter-embeddinggemma-2-for-text-image-audio-and-vi/
- Seen in: DAIR.AI (Elvis Saravia)
- Change: capability (New capability)
- Applies to: EmbeddingGemma 2, Hugging Face, LiteRT

## What changed

Google released EmbeddingGemma 2, an open 744M-parameter model that maps text, images, audio, video and mixed inputs into one vector space, available on Hugging Face and LiteRT.

Google researchers have released EmbeddingGemma 2, an open model with 744M parameters. It maps text, images, audio, video and mixed inputs into one vector space, so content in one medium can be compared directly with content in another. The arXiv listing, dated October 27, 2026, says the models are available on Hugging Face and LiteRT. An embedding model turns an input into a list of numbers, and inputs with similar meaning land close together. That is what makes search and retrieval work across media.

The model is assembled from parts. A 167M vision encoder and a 305M audio encoder feed a 271M bidirectional transformer. Because the encoders are separate, a team that only needs text can load the 271M configuration. The vision-language and audio-language configurations come to 439M and 577M. The paper says the model builds on the Gemma 4 family and learns from Gemini Embedding 2. Training runs in three stages that separate text-only from multimodal training and distillation from contrastive objectives.

The abstract says the model "establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger." The authors report results on seven benchmark suites covering text, code, image, video, document, speech and sound. Among the figures given, MTEB Multilingual v2 reaches 61.36, and MTEB Code rises from 68.76 to 78.68 over the predecessor. MTEB English v2 falls from 69.67 to 68.46. The frontier language does not mention that regression.

The deployment features show where the design choices lie. Teams can quantize the weights to int4, which the paper's figures show costs about 0.4 points on MTEB Multilingual. They can also lower the vision token budget, meaning fewer tokens represent each image. They can truncate the embedding dimensions, which shortens each stored vector. The authors frame these options as flexibility for resource-constrained environments, and the abstract presents them alongside selective encoder loading.

This report gives a handful of benchmark scores without the complete per-suite tables or the comparison models behind them. So the claim of beating models two to twelve times larger cannot be checked against the specific numbers here. The English regression and the multilingual score appear without the baselines that would make them easy to judge. Readers evaluating the model for their own workloads would need the full results in the paper.

## Why it matters

Companies running retrieval over mixed media on devices or in latency-sensitive systems get an open option that gives up little accuracy for a much smaller model. Modular loading lets teams deploy only the encoders they need.

## What you can do

Benchmark the int4 or truncated-dimension versions on your own retrieval data before choosing an embedding stack.

## Original source

- [(2610.37364) EmbeddingGemma 2: Unified and Efficient Multimodal Representations](https://arxiv.org/abs/2610.37364)
- [EmbeddingGemma 2: Unified and Efficient Multimodal Representations | DAIR.AI Academy | DAIR.AI Academy](https://academy.dair.ai/papers/embeddinggemma-2-unified-and-efficient-multimodal-representations-2610.37364)

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Google presents open 744M-parameter EmbeddingGemma 2 for text, image, audio and video", 2026-11-01, https://ai-news.ikangai.com/story/google-presents-open-744m-parameter-embeddinggemma-2-for-text-image-audio-and-vi/
