Google details EmbeddingGemma 2, an open multimodal embedding built for on-device use
Google on Tuesday pointed developers to a new technical report for EmbeddingGemma 2, offering a fuller look at an open multimodal embedding model that the company says is small enough to run on devices rather than only in the cloud.
In a short blog post published Nov. 17, Google framed the paper as a deeper explanation of a model it first announced on Oct. 6. “We released EmbeddingGemma 2, which expands our text-only open embedding model with native multimodal understanding,” the company wrote. “It maps text, images, audio, and video, or any combination of them, into a single embedding space, all while compressing the model backbone into a more compact profile.”
Embedding models convert content into numerical representations, or vectors, so software can compare meaning across documents, images or other media. According to Google, EmbeddingGemma 2 is an open, natively multimodal system that places text, code, images, audio, video and mixed inputs into one shared vector space. The model is released under the Apache 2.0 license, and Google says weights are available publicly through Hugging Face and Kaggle.
The technical report, titled “EmbeddingGemma 2: Unified and Efficient Multimodal Representations,” is posted on arXiv as paper 2610.37364. ArXiv shows a first version submitted Oct. 27, 2026, and a revised version posted Nov. 9. The paper lists Henrique Schechter Vera and Sahil Dua as lead authors among a larger Google DeepMind author group.
Google says the model has 740 million parameters and is built on the Gemma 4 architecture. Its official model card breaks that total into a 270 million-parameter text model — including a 130 million-parameter transformer and a 140 million-parameter embedder — plus a 170 million-parameter vision encoder and a 300 million-parameter audio encoder.
The native output is a 768-dimensional embedding, meaning each item is represented as a list of 768 numbers. Google also says the model supports Matryoshka Representation Learning, a technique that lets developers truncate those embeddings to 512, 256 or 128 dimensions to save storage, with claimed reductions of up to sixfold. The model card says the system supports an 8,192-token context window for combined multimodal inputs.
A key part of Google’s pitch is local use. In its Oct. 6 launch post, the company said the model is “optimal for on-device inference.” The model card gives example memory figures on a Pixel 11 Pro of about 191 megabytes of active RAM for text-only weights and about 567 megabytes for the full multimodal model, with quantization, a method for shrinking model size.
In the paper, the authors argue that EmbeddingGemma 2 “establishes a new Pareto frontier,” their term for a stronger trade-off between quality and efficiency, and say it outperforms specialist models and models two to 12 times larger on their benchmarks. That is the authors’ characterization, not an independent verification. One concrete example from Google’s reported results: on the MTEB code benchmark, the company says EmbeddingGemma 2 reached an NDCG@10 score of 78.68, compared with 68.76 for EmbeddingGemma 1, a 9.92-point absolute gain.
The release adds detail to a product line Google began in September 2025 with EmbeddingGemma, a text-only embedding model. EmbeddingGemma 2 expands that approach into native multimodal retrieval, while keeping the model open and commercially permissive. The Oct. 6 announcement introduced the model; Tuesday’s post mainly gives developers the architectural and evaluation detail behind it.