Google has open-sourced EmbeddingGemma 2, its multimodal embedding model, straight to Hugging Face: 740M parameters under Apache 2.0, built to keep retrieval workloads on phones instead of shipping data to the cloud for vectorization.

One Model, Five Modalities

Built on the Gemma 4 architecture, EmbeddingGemma 2 maps text, code, images, audio, and video into a single shared embedding space. The official blog gives concrete use cases: find a specific video clip from a voice memo, or search hours of audio recordings with a text query — all processed locally by one natively multimodal model. Its predecessor, EmbeddingGemma, was text-only; Google says it has passed 20 million downloads, mostly powering on-device search and privacy-first RAG pipelines. Multimodal support is exactly the piece this generation adds.

Modular by Design: Pay Memory Only for What You Use

The most pragmatic design choice is modularity: text-only workloads need as little as 270M parameters, with optional vision (170M) and audio (300M) encoders for full multimodal support. Using Matryoshka Representation Learning (MRL), output vectors can be dynamically truncated from 768 dimensions down to 512, 256, or 128 — Google claims up to 6x storage reduction for local vector databases. The context window is 8K tokens, 4x larger than EmbeddingGemma 1, fitting about 5.5 minutes of audio, 29 images, or 58 video frames. With quantization on a Google Pixel 11 Pro, the model needs as little as ~191MB active RAM for text-only weights and ~567MB for the full multimodal stack — single-device territory.

Benchmarks and Ecosystem

Official numbers show MTEB Code improving from 68.76 to 78.68, a 9.92-point jump over the previous generation. Google claims best-in-class scores among sub-1B multimodal embedders, outperforming some specialist models more than twice its size — note these are all vendor-reported figures. The ecosystem support is thorough: weights are on Hugging Face and Kaggle, with llama.cpp, Ollama, MLX, vLLM, and SGLang all supported from day one. It also shares its text tokenizer and audio encoder with Gemma 4, lowering the combined memory footprint when both run together — the intent of a "Gemma 4 for generation + EmbeddingGemma 2 for retrieval" on-device RAG stack is hard to miss.

So What

In the on-device AI narrative, generative models get the spotlight, but retrieval is the harder problem: without local embeddings, privacy-first RAG still has to send data to the cloud for vectorization. Google has now filled that gap — and with Apache 2.0 rather than a restricted license, a real concession to edge developers. The question worth watching is whether other edge ecosystems will follow with comparable open embedding models. The retrieval layer is becoming the new front line of on-device AI competition.

(Source: Google's official blog)