Google releases EmbeddingGemma 2 for multimodal search on edge devices

Google announced on October 6, 2026 the release of EmbeddingGemma 2, an open‑source multimodal embedding model licensed under Apache 2.0. The model is positioned as a lightweight solution for on‑device retrieval, routing and similarity search across text, code, images, video and audio. By making the weights publicly available on Hugging Face and Kaggle, Google invites developers to run the model directly on edge hardware without relying on cloud APIs. The announcement emphasizes that the model is not intended for generative tasks but for offline embedding generation that can power local search engines, recommendation pipelines and content‑based filtering.

A lightweight multimodal embedding model

EmbeddingGemma 2 contains 740 million parameters that jointly encode five data modalities—plain text, source code, still images, video frames and audio snippets—into a single embedding space. The unified space enables a single similarity search index to handle heterogeneous queries, for example matching a spoken description to relevant code snippets or retrieving video frames that correspond to a textual prompt. Google describes the architecture as a transformer‑based encoder that processes each modality through dedicated front‑ends before projecting the representations into a common 768‑dimensional vector. The model’s open‑source nature allows researchers to inspect the layers, fine‑tune on domain data, or replace modality‑specific components.

For pure text workloads Google reports a reduced configuration that uses only 270 million parameters. Developers who need visual understanding can attach an optional 170‑million‑parameter vision encoder, while audio processing adds a further 300‑million‑parameter encoder. The modular design means that a developer can run a text‑only instance on very constrained devices and only load the additional encoders when multimodal input is required. Google states that the optional encoders are compatible with the same embedding head, preserving the unified vector format across all supported modalities.

Matryoshka Representation Learning for storage efficiency

A key innovation highlighted by Google is Matryoshka Representation Learning, a technique that lets developers compress the default 768‑dimensional vectors to 512, 256 or 128 dimensions without retraining the model. The method works by learning nested sub‑spaces that retain most of the similarity information while discarding redundant dimensions. According to Google’s measurements, moving to 128 dimensions can shrink the storage footprint by up to six times, which translates into lower memory consumption for large index tables on edge devices. The compressed vectors remain compatible with standard similarity metrics such as cosine similarity, enabling drop‑in replacement in existing pipelines.

Edge performance on Pixel 11 Pro

Google provides device‑level benchmarks for the Pixel 11 Pro, the flagship phone released earlier in 2026. A quantized text‑only version of EmbeddingGemma 2 occupies roughly 191 MB of active RAM, while the full multimodal stack—text, vision and audio encoders combined—uses about 567 MB. The model supports an 8,000‑token context window for text, up to 5.5 minutes of audio, 29 images or 58 video frames per inference pass. These figures demonstrate that a single high‑end smartphone can host the entire model and perform real‑time embedding generation for mixed‑media queries without off‑device assistance.

In the MTEB (Massive Text Embedding Benchmark) suite, Google reports that EmbeddingGemma 2 improves the Code sub‑benchmark score from 68.76 to 78.68. The increase reflects the model’s ability to capture syntactic and semantic patterns in programming languages, a capability that benefits code search and duplicate detection. Google positions the result as evidence that a 740‑million‑parameter open model can outperform larger proprietary embeddings on specific tasks, though the benchmark is a vendor‑provided measurement and may not capture all real‑world workloads.

The model weights are hosted publicly on both Hugging Face and Kaggle, and Google notes compatibility with common runtimes such as TensorFlow, PyTorch and ONNX. By providing conversion scripts and example inference pipelines, Google aims to lower the barrier for developers to integrate EmbeddingGemma 2 into mobile apps, embedded systems or edge servers. The open license (Apache 2.0) permits commercial use, modification and redistribution, which aligns with Google’s broader strategy of fostering an ecosystem of community‑driven AI tools that run locally.

Google cautions that the performance numbers are vendor measurements taken on a specific hardware configuration and may vary on other edge platforms. The model is optimized for retrieval and routing, not for generative text or content creation, so developers should not expect it to replace large language models in chat or summarisation scenarios. Additionally, the storage and RAM requirements, while modest for a flagship phone, could still be prohibitive for low‑end IoT devices lacking several hundred megabytes of free memory.

For developers building private, offline search solutions, EmbeddingGemma 2 offers a concrete path to unify multimodal indexing without sending data to the cloud. Applications such as on‑device personal assistants, secure document management systems, or localized recommendation engines can now embed text, code, images, video and audio in a single vector space and perform similarity matching locally. The ability to shrink vectors with Matryoshka further reduces the cost of maintaining large indexes, making it feasible to store millions of items on edge storage media.

With EmbeddingGemma 2, Google changes the baseline for what can be run on‑device: a single open model now handles five modalities, fits within the memory envelope of a modern smartphone, and delivers benchmark‑level code understanding. The release signals a shift toward more capable, privacy‑preserving AI that lives at the edge, and it invites the developer community to experiment with multimodal retrieval without relying on heavyweight cloud services.