Google Releases EmbeddingGemma 2 With Multimodal Capabilities for On-Device Use
Google released EmbeddingGemma 2, an open-weight, multimodal embedding model designed to run on consumer hardware such as smartphones and laptops. The model expands the EmbeddingGemma line beyond text handling, which was the sole capability of the first version introduced in September 2025. Built on the Gemma 4 architecture released in April, the new model allows developers to index and search text, code, images, video, and audio within a unified 768-dimensional embedding space.
The model scales from 270 million parameters for text and code up to 740 million parameters for full multimodal support, letting developers load only the encoders required for their specific applications. According to Google DeepMind research engineers Sahil Dua and Henrique Schechter Vera, developers have downloaded the original EmbeddingGemma model more than 20 million times. The updated model is available on Hugging Face and Google’s Kaggle under an Apache 2.0 license permitting commercial use, and its weights work with open-source serving tools including vLLM, llama.cpp, and Ollama.
Hardware Memory Footprint and Vector Storage Reductions
Running on a quantized build on a Google Pixel 11 Pro, the 270 million-parameter text core uses about 191 megabytes of memory. Because applications must also store the vectors generated by the model, storing a million 768-dimensional vectors at bfloat16 precision takes roughly 1.5 gigabytes of memory. To mitigate storage demands, EmbeddingGemma 2 supports Matryoshka Representation Learning, a training technique that lets developers truncate vectors down to shorter lengths.
Truncating vectors to 128 dimensions reduces storage requirements by six times down to 250 megabytes, though it introduces a retrieval quality trade-off. At 128 dimensions, the model retains approximately 90 percent of text quality, while performance for image, video, and speech retrieval drops to around 75 percent. Google’s developer guide notes that truncating vectors to 256 dimensions allows image, video, and speech retrieval to keep about 95 percent of their full quality.
Performance Benchmarks Across Code Search and Multilingual Tasks
Code search showed the most significant benchmark gains for the updated model. EmbeddingGemma 2 scored 78.68 on the code section of the Massive Text Embedding Benchmark, outperforming the initial release by nearly 10 points. Google states that this performance makes the model well-suited for local codebase indexing and agentic code search. Meanwhile, multilingual text scores barely moved.
Reddit user Distinct-Target7503 commented that the model’s ability to embed images and audio in a shared semantic space is impressive, though they noted it does not represent a massive improvement for text-only tasks. Another Reddit user, Open-Adhesiveness-86, cautioned developers that task prefixes are mandatory and that truncating vectors requires slicing before applying L2-normalization to prevent off-norm vectors.
Integration Plans Within Google Ecosystems
Google plans to add the model weights to the Model Garden in the Gemini Enterprise Agent Platform soon. Because Gemma 4 and EmbeddingGemma 2 share an audio encoder and a text tokenizer, deploying both models together for an on-device retrieval-augmented generation workflow consumes less memory than pairing two distinct models. Google’s AI Edge Foresight meeting application for Mac already runs the pair together, and the AI Edge Gallery demo app features a Video Moments Finder that locates specific scenes inside videos using typed or spoken queries.
Keep reading