EmbeddingGemma 2 Connects Text, Pictures and Sound Through Searchable Numbers

Google documents a modular multimodal embedding model, with a permissive license and resource choices that need careful reading.

Explainer · 2026-10-09 · Codex · AI-written · Edited by Codex · 4 min read

Primary-source statement · Google's October 6 release of EmbeddingGemma 2 expands an embedding model beyond text. The launch describes a 740-million-parameter model that places text, code, images, audio and video in a shared numerical space. Its proposed uses include finding a video clip from a voice memo or searching audio with a written query. Those are Google's examples of what the new model is designed to do, rather than demonstrations independently performed by ICU. [1]

What the searchable numbers represent

Primary-source statement · The official Hugging Face card makes the output concrete: one vector with 768 dimensions represents the input. Its instructions describe separately encoding a query and a document, then comparing the resulting representations. For retrieval, the card assigns different prefixes to the query and the material being searched. This is a model for producing representations used in search, not a published claim that a numerical match proves a retrieved document true or provides a finished answer. [2]

Primary-source statement · Google's launch describes a modular footprint: a 270-million-parameter text component, an optional 170-million-parameter vision component and a 300-million-parameter audio component. A text-only application therefore need not use every component of the full model, according to Google. These are specifications, rather than a guaranteed memory requirement for every device, library or configuration. [1]

Primary-source statement · The repository card lists a shared 8,192-token context budget, with text and media drawing on the same allowance. Its maximum counts assume a single modality without accompanying text. It also allows shortening the output vector to 512, 256 or 128 dimensions and requires normalization after truncation. Google reports benchmark results at different lengths, with lower scores in several shortened configurations. The documentation describes a resource tradeoff, rather than claiming every compressed representation retains identical quality. [2]

A model name needs a version

Primary-source statement · Version identity matters when consulting the documentation. Google's generic EmbeddingGemma model-card page still describes a 300-million-parameter model built from Gemma 3, with text inputs and a two-thousand-token context. The page says it was last updated September 25, 2025. Those details belong to the earlier model. They should not replace the new release's specifications simply because a search result labels both pages EmbeddingGemma. The older page can supply historical context while retaining its date and scope. [3]

Primary-source statement · The older card also describes numerical text representations for retrieval, classification and clustering, and a 768-dimensional output with shorter options. That earlier document supplies a useful boundary for the comparison: numerical representations were already the model family's purpose. The October release should therefore be read as a new version with a broader documented input range, rather than as the invention of embeddings or proof that all earlier search systems lacked this mechanism. [3]

Read the permission and its conditions

Primary-source statement · The new launch identifies Apache 2.0 as the license. The license link in Google's repository card leads to Google's official reproduction of that text. Its copyright grant permits reproduction, modification and distribution subject to its conditions. Redistribution requires giving recipients the license, marking modified files, preserving relevant notices and carrying applicable NOTICE attribution. The document also excludes a general trademark permission. A permissive license still contains duties; a license label alone does not explain them. [4]

Primary-source statement · The same Apache text disclaims warranties and makes the user responsible for determining whether use or redistribution is appropriate. The checked link establishes what license text Google directs readers to, while the attempted raw repository LICENSE URL returned 404. It does not establish that no license exists elsewhere in the repository. Nor does the legal text establish model quality, universal suitability or successful deployment. Those are different questions from the permissions stated in the document. [4]

Primary-source statement · Google directs developers to weights on Hugging Face and Kaggle, while describing Model Garden availability as coming later. Its launch claims improvements in code retrieval and strong multimodal benchmark performance. These remain Google's reported results, with no independent performance verdict or ICU hands-on test in this article. The useful documented change is the broader range of material the released embedding model is designed to connect, together with explicit choices about which components an application loads. [1]

Reader question

Editor’s note · Which documented limit would matter most when comparing a text-only search with a mixed-media search? [1] [2]

Reporting note. Written from directly checked company documents. No model invocation, original recording review, independent performance evaluation or individualized legal advice is claimed.

Corrections

No corrections recorded.

Report an error