Google's EmbeddingGemma 2 fits multimodal search into 567MB of phone RAM

Google DeepMind launched EmbeddingGemma 2, expanding beyond text to unify code, images, video and audio in a shared embedding space.

EmbeddingGemma 2 requires as little as 270M parameters for text-only workloads, with optional vision (170M) and audio (300M) encoders for full multimodal support.

Using Matryoshka Representation Learning, developers can truncate output vectors from 768 dimensions down to 512, 256 or 128, providing up to 6x storage reduction for local vector databases and memory usage.

The RAM figures are Google's measurements, not independently verified device tests.

Different kinds of input are mapped into the same 768-dimensional space, so a text query can be compared with an image, video frame or audio clip.

The model card sets a shared 8,192-token context budget and estimates that with one modality it covers about 5.5 minutes of audio, 29 images or 58 video frames; mixing media consumes the same budget, and video is sampled at one frame per second by default.

EmbeddingGemma 2 improves MTEB Code from 68.76 to 78.68, a 9.92-point gain, which Google says makes it well-suited for local codebase indexing, semantic code search and coding agent retrieval.

EmbeddingGemma 2 has 740 million parameters, is built on the Gemma 4 architecture and is released under a commercially permissive Apache 2.0 license.

On the multilingual MTEB benchmark, the listed scores are 61.36 for EmbeddingGemma 2 and 61.15 for the first EmbeddingGemma.

The benchmark figures are published by Google and do not establish how the model performs on any particular private collection of photos, recordings, documents or code.

Retrieval often depends on a chain of separate models; Google's AI Edge post says EmbeddingGemma 2 can replace parts of that chain, such as image captioning, speech-to-text and text embedding, with local vector search, and that the searches run on-device without an internet connection.

EmbeddingGemma 2 shares Gemma 4's text tokenizer and audio encoder, so developers can run both models together in a unified pipeline with a lower combined total memory footprint.

The model can be served with transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio.

Weights are available through Hugging Face, with Kaggle also listed; Google says support through Android's ML Kit is planned for the coming weeks.

The first EmbeddingGemma had more than 20 million downloads.

Google documents Gemini Embedding 2 as a multimodal model offered through the Gemini API; EmbeddingGemma 2 is a local counterpart that puts weights in developers' hands, and developers weigh local hardware and indexing work against a hosted service.

The up-to-6x storage reduction equals the ratio of 768 to 128 dimensions.

Truncating to 128 dimensions costs 13.36 points on the multimodal benchmark, about 23% of the full score.

Truncating to 128 dimensions costs 7.27 points on the code benchmark, about 9% of the full score.

The modular components sum to the stated total: 270M + 170M + 300M = 740M parameters.

Loading the vision and audio encoders adds about 376MB of active RAM over text-only use on Google's Pixel 11 Pro figures.

The relative quality loss from 128-dimension truncation is about 2.5 times larger for multimodal retrieval than for code.

At the default one frame per second, the 58-frame window covers about 58 seconds of video.

An hour of audio needs at least 11 separate context windows at about 5.5 minutes each.

Multilingual MTEB moved 0.21 points between versions, against a 9.92-point gain on MTEB Code.