EmbeddingGemma 2: Multimodal Embedding Model Released

👁 114 reads

Google DeepMind’s EmbeddingGemma 2 is a 740-million parameter embedding model that maps text, code, images, video and audio into a single shared vector space - all while running directly on your device.

Google DeepMind released EmbeddingGemma 2 on October 6, 2026. This is the first version capable of full multimodality while maintaining a size suitable for on-device operation.

The model is built on the Gemma 4 architecture and released under the Apache 2.0 license, allowing commercial use without additional fees. According to DeepMind, it’s a direct derivative of the same technology underlying the full-scale Gemini Embedding models.

Modular Architecture and Memory Efficiency

EmbeddingGemma 2 has a modular construction: the text core requires only 270 million parameters, and you can optionally attach a vision encoder (170M parameters) or audio encoder (300M) to achieve full multimodal capability. On a Google Pixel 11 Pro smartphone, the model needs just about 191 MB of active RAM after quantization for text-only mode and about 567 MB for full multimodal mode - realistic figures even for average mid-range phones.

Additional disk space savings come from Matryoshka Representation Learning (MRL), which lets developers dynamically truncate output vectors from 768 to 512, 256 or 128 dimensions. According to DeepMind, this provides up to six-fold reduction in local vector database size.

Context Window Four Times Larger Than Predecessor

EmbeddingGemma 2 supports an 8K token context window - four times more than the first version. In practice, this means the ability to process in a single call up to 5.5 minutes of audio recording, 29 images, 58 video frames or interleaved combinations of these - all locally, without sending data to the cloud.

In the MTEB Code benchmark, the model scored 78.68 points versus 68.76 in the previous version, an improvement of 9.92 points. DeepMind emphasizes that EmbeddingGemma 2 outperforms some specialized models with more than twice as many parameters in tasks involving images, video, documents and audio.

Previous Version Downloaded Over 20 Million Times

The success of the first edition was surprising according to DeepMind: EmbeddingGemma 1 exceeded 20 million downloads and made its way into on-device search applications and private RAG (retrieval augmented generation) pipelines. This popularity prompted the team to extend the model’s capabilities beyond text alone.

Practical applications of EmbeddingGemma 2 include searching for a specific video clip based on a voice note, searching through hours of audio recordings using a text query, or semantically indexing local code repositories - all without any data leaving your device.

For mobile app developers looking for a viable alternative to cloud-based embedding APIs, this is a proposition that’s hard to overlook.

// read moreEmbeddings and Vector Databases Explained (Polish Test)GuidesRouting Queries to Multiple Embedding Models Without CurseState of AIUse LLM Hallucination to Tag Content, Then Anchor It With EmbeddingsGuides

Kinetyra