Google Releases EmbeddingGemma 2: Open Embedding Model Expands to Text, Images, Audio and Video, With a Public Technical Report
Google says it has released EmbeddingGemma 2, extending its text-only open embedding model to native multimodality, and has published a technical report. This article covers how it differs from the previous version, what the report includes, and what Google has not yet stated, such as license, model size and platforms.
News date: About 4 min read
Contents
In an official blog post dated November 17, 2026, Google said it has released EmbeddingGemma 2 and has made a technical report available to readers. An embedding model is an AI model that turns content into a series of numbers called a vector. Similar content gets similar numbers, which makes embeddings useful for search and matching. Multimodal means handling more than one type of data. According to Google, the model natively supports text, images, audio and video. The post does not mention a staged rollout, a preview or regional restrictions, so check availability on Google's official pages.
How does EmbeddingGemma 2 differ from earlier versions?
According to Google, EmbeddingGemma 2 builds on the earlier text-only open embedding model by adding native multimodal understanding. It can map text, images, audio and video, or any combination of them, into the same embedding space. This means different types of content are converted into one shared numerical representation. Google also says the model backbone (the core structure of the model) has been compressed into a more compact form.
| Aspect | Earlier EmbeddingGemma | EmbeddingGemma 2 (per Google) |
|---|---|---|
| Supported data types | Text only | Text, images, audio, video, or any combination |
| Embedding space | Described only as a text embedding model | A single, unified embedding space |
| Model backbone | Not mentioned | Compressed into a more compact form |
| Technical foundation | Not mentioned | Built on technology from Gemma 4 and Gemini Embedding 2 |
What does the technical report cover?
Google says readers can find the training strategy, evaluation methods and experimental results in the official technical report. Google also notes that the innovations in the model recipe build on the latest technology from Gemma 4 and Gemini Embedding 2. The goal is to learn an expressive, unified vector space. The post also includes a five-minute video explaining how the model works and how it was developed.
What does a shared space mean for developers?
As Google describes it, different media types are converted into the same numerical representation. They can therefore be compared within one vector space, without a separate model for each type. Google's post gives no figures for actual accuracy, speed or cost. Judging these requires reading the technical report or evaluating the model yourself.
FAQ
Is EmbeddingGemma 2 officially launched now?
Google says in the post, "We released EmbeddingGemma 2," which indicates it has been released. However, the post does not say whether it is a preview or beta, or whether there are plan or regional restrictions. Check the official release page for these details.
Where can I find the training and evaluation details?
Google says the training strategy, evaluation methods and experimental results are included in the official technical report. The blog post also provides a five-minute explainer video.
Is it available in my country or region?
Google's post does not mention any regional availability or restrictions, so this cannot be confirmed yet. Rely on Google's official information.
Browse the latest news in this topic
Found this guide helpful? Share it with someone who needs it
NVIDIA and Microsoft at Ignite: Vera Rubin NVL72 in Preview on Azure, Nemotron 3.5 in Foundry, Windows 365 Direct Stream in Limited Preview
NVIDIA says it has expanded AI collaboration with Microsoft from the Azure cloud to the edge and Windows, including a Vera Rubin NVL72 preview on Azure, Nemotron 3.5 models in Foundry and a limited preview of Windows 365 Direct Stream.
Google Quantum AI Makes the Cover of Science: Quantum Contextuality Holds at Up to 105 Qubits, With a 98.3% Win Rate in a Coordination Game
Google: ARMORY CASTLE Exploited VMware vCenter Zero-Day CVE-2026-123890, Affecting vCenter 7.x and 8.x; Broadcom Has Released a Patch
Cloudflare: Election-Day Traffic to 2026 U.S. Midterm Election Sites Surged 363%, While Attackers Shifted from DDoS to Exploiting Vulnerabilities
Sources
Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.