TellWorldAI World Archive
Ask中文

Home›Models

Model

EmbeddingGemma 2

EmbeddingGemma 2 is Google's embedding model that extends the text-only EmbeddingGemma to text, code, images, video and audio in one 768-dimensional space. It is released under the Apache 2.0 license.

Modality Text, Image, Audio, VideoOpen weights YesContext (K tokens) 8 Era Generative AI for everyone

Parameters
740M
License
Apache-2.0
Access
Download
Release date
Developer
Google DeepMind

Background

An embedding model turns a piece of content into a list of numbers called a vector. Content with similar meaning ends up with vectors that sit close together, which makes embeddings useful for search, retrieval and matching. EmbeddingGemma is Google's line of embedding models, and the first version worked on text only.[1]

What it is

On 17 November 2026 Google announced EmbeddingGemma 2, which adds native multimodal understanding to the text-only model.[1] It maps text, including code, images, video and audio, alone or in combination, into a single 768-dimensional vector space.[2] Because all of these inputs share one space, a text query and an image or video can be compared directly, using the same kind of distance calculation.

Google says the model builds on the architectural and capability advances of Gemma 4, the company's family of open models.[2]

Key details

The model has 740M total parameters: a 270M text model plus modular vision (170M) and audio (300M) encoders.[2] An encoder is the part of the network that processes one type of input. Because the vision and audio encoders can be loaded selectively, developers can load only the modalities they need.[2]

The supported Matryoshka truncation dimensions are 128, 256, 512 and 768.[2] Truncation means a vector can be shortened to a smaller size. Shorter vectors take less storage and are quicker to compare, at some cost in precision.

The context window is 8,192 tokens, which sets how much text a single input can cover.[2] A token is a small unit of text, often a word or part of one.

Google advises running inference in bfloat16 or float32 and not float16, because the model's activation range exceeds float16's dynamic range.[2] Inference here means running the model to produce results, and the number format determines which values can be stored without overflow.

Licensing

The license is Apache 2.0.[2] Under that license, the model can be used, modified and redistributed under permissive terms, which matters for developers who want to run it in their own systems.

Why it matters

Embedding models are a quiet building block. They do not write answers themselves, but they decide which documents, images or clips a search system finds. Extending one model to several kinds of input, with the option to load only the parts needed, lets developers build retrieval over mixed content without assembling separate systems for each type.

20261

  1. ReleaseGoogle releases EmbeddingGemma 2 with multimodal embeddings

    On 17 November 2026 Google released EmbeddingGemma 2, which adds multimodal embeddings to the text-only model. Google reports a code benchmark score of 78.68, up from 68.76 for the first version.

Developed by1

Google DeepMind

Builds on1

Gemma 4

Notes and sources

  1. Read the EmbeddingGemma 2 technical reporthttps://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2-technical-report/
  2. EmbeddingGemma 2 model cardhttps://ai.google.dev/gemma/docs/embeddinggemma/model_card_2

Cite this page: EmbeddingGemma 2. TellWorld, https://www.tellworld.net/models/embeddinggemma-2. Updated 2026-11-18