Google EmbeddingGemma 2

Text, source code, images, video frames, and audio all map into the same 768-dimension space. You can query an audio archive with a written sentence, or find a screenshot using a description of what it showed. Cross-modal comparison works because the vectors share a coordinate system, not because separate models were stitched together afterward. Most multimodal embedding model setups encode each format in its own space and translate between them, which loses detail at the boundary.

The full model is 740M parameters, split into a 270M text and code backbone, a 170M vision encoder, and a 300M audio encoder. Ship the text path alone and the download and memory cost drop sharply. Add vision or audio later, and previously computed text vectors stay valid because the space doesn't change.

Matryoshka Representation Learning packs the most useful information into the front of each vector. You can keep all 768 dimensions, or cut the output to 512, 256, or 128 for local storage savings of up to 6x. Lower dimensions cost some retrieval accuracy, and the loss shows up sooner on image and audio than on text, so test your own queries before committing.

The context window is four times larger than the first EmbeddingGemma's. That translates to roughly 5.5 minutes of audio, 29 images, or 58 video frames in a single input, including interleaved combinations. A slide deck with narration can be encoded as one item instead of a pile of separate chunks.

With quantization, text-only weights take about 191MB of active RAM on a Pixel 11 Pro. The full multimodal model needs around 567MB. Those numbers make the difference between an app that ships and one that gets rejected for memory use, especially on mid-range Android hardware.

On MTEB Code, the score moved from 68.76 to 78.68, a 9.92 point gain over the previous version. The model handles local codebase indexing and retrieval for coding agents, where a query often describes intent rather than naming a symbol. Google reports it also beats some specialist models more than twice its size on image, video, document, and audio tasks.

Official and community integrations cover sentence-transformers, transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, transformers.js, and WebGPU. Qdrant published a walkthrough for storing the vectors, and Unsloth documents fine-tuning. You aren't locked into a single inference stack or a Google-only deployment path.

EmbeddingGemma 2 shares its text tokenizer and audio encoder with Gemma 4. Running both in one pipeline avoids loading duplicate components, which lowers the combined memory footprint. The retrieval step and the generation step can then happen on the same laptop or phone without a network call.