EmbeddingGemma 2 Brings Multimodal Search to Devices

Google DeepMind released an open-weights model that maps text, audio, images, and video into a shared coordinate space on local hardware.

Searching private files across different media formats usually requires a cumbersome chain of specialized models. On October 6, 2026, Google DeepMind released EmbeddingGemma 2 under the Apache 2.0 license to streamline that process. Built on the Gemma 4 family, the open-weights model generates vector representations directly across text, code, images, audio, and video while running entirely on consumer hardware.

EmbeddingGemma 2 Maps Five Modalities Into One Space

Traditional search pipelines handle mixed media through sequential conversions. Querying video recordings with plain text often means running automated speech recognition to transcribe dialogue, extracting frame captions with a vision system, and embedding the generated text strings into a vector database.

EmbeddingGemma 2 discards those intermediate processing layers. It projects written text, source code, photos, audio clips, and video frames into a shared 768-dimensional mathematical space. Because the vectors share coordinates, a single text prompt retrieves matching audio timestamps or exact video moments without prior transcription.

Modular Design and Memory Limits

The architecture tops out at 740 million parameters, but developers do not need to load the complete multimodal stack. DeepMind split the model into modular tiers to fit smaller devices:

Text and code base: 270 million parameters.
Text and vision: 440 million parameters.
Text and audio: 570 million parameters.
Full multimodal: 740 million parameters.

The context window grew from 2,048 tokens in the original model to 8,192 tokens. According to DeepMind technical documentation, this window accommodates long prose, 29 static photos, 58 video frames, or up to 5.5 minutes of raw audio in a single pass. In internal company benchmarks run on a Google Pixel 11 Pro with mobile quantization, the text module required 191 MB of active RAM, while the full 740M multimodal package consumed roughly 567 MB.

Matryoshka Truncation and Benchmark Scores

To conserve storage when indexing millions of records locally, the model supports Matryoshka Representation Learning (MRL). Teams can truncate the default 768-dimensional outputs down to 512, 256, or 128 dimensions at generation time, cutting disk and memory footprints by up to six times with modest accuracy trade-offs.

In evaluation data published in DeepMind’s Model Card, EmbeddingGemma 2 posted 78.68 points on MTEB Code, a 9.92-point gain (about 14%) over version 1. The company also reported first place among sub-1-billion parameter models on the Massive Audio Embedding Benchmark (MAEB) and confirmed multilingual support spanning more than 100 languages. These figures reflect company-provided metrics; independent community benchmarks will test how well those numbers hold up in everyday production deployments.

Offline RAG in Open Runtimes

The first release of EmbeddingGemma logged more than 20 million downloads but stayed confined to short text snippets. EmbeddingGemma 2 targets private, on-device retrieval-augmented generation (RAG) where sensitive recordings, local documents, and media archives stay off commercial cloud endpoints.

Third-party developers packaged the weights immediately. Community maintainers at Unsloth AI published quantized GGUF builds, and the model integrates out of the box with open-source runtimes including llama.cpp, Ollama, MLX, LiteRT, vLLM, SGLang, Qdrant, and sentence-transformers.