DeepMind Debuts EmbeddingGemma 2, Mapping Five Modalities Into One Space

Google DeepMind has unveiled its latest AI model, EmbeddingGemma 2, which maps five different modalities - text, code, images, video, and audio - into a single 768-dimensional embedding space. This open model boasts an impressive 740 million parameters and is designed to run on consumer hardware. The release of EmbeddingGemma 2 follows the success of its predecessor, which has seen over 20 million downloads since its introduction in 2025 as a lightweight option for text embeddings on consumer devices.

The new model was unveiled by Google DeepMind research engineers Sahil Dua and Henrique Schechter Vera in a post on Google’s official blog. According to the company, EmbeddingGemma 2 is the most capable model for on-device multimodal embeddings and is built from the same technology as its Gemini Embedding models. The capabilities of this model are vast, with examples including retrieving specific video clips with voice memos or querying hours of audio recordings using text.

EmbeddingGemma 2 is based on the Gemma 4 architecture, which combines a 270-million-parameter text model with a 170-million-parameter vision encoder and a 300-million-parameter audio encoder. This modular design allows developers to selectively load parameters for different modalities, such as loading only the text and code configuration or the full multimodal setup from the same checkpoint. The shared vector space across all modalities enables efficient processing of multiple inputs simultaneously.

The model card lists a comprehensive set of features, including a 24-layer architecture with grouped-query and multi-query attention, a 262,144-token vocabulary, mean pooling, and a projection layer from 512 to 768 dimensions. All modalities share an 8,192-token context window, which is four times larger than its predecessor’s. Each modality consumes this budget at fixed rates: 280 tokens per image, 140 tokens per video frame, and 25 tokens per second of audio.

The model card also reports benchmark results for EmbeddingGemma 2, showing that it matches the multilingual text performance of its predecessor while raising its MTEB Code score by 9.92 points to 78.68. The full-precision results list impressive scores across various tasks, including 61.36 on MTEB multilingual (v2), 64.64 on MIEB lite, and 69.54 on MSEB sound retrieval. Google claims that EmbeddingGemma 2 outperforms some specialist models more than twice its size.

One of the key features of EmbeddingGemma 2 is Matryoshka Representation Learning, which allows developers to truncate output vectors from the native 768 dimensions down to 512, 256, or 128. This reduces vector storage requirements by up to six times and maintains quality with minimal impact down to 256 dimensions. However, using 128 dimensions may compromise performance on multimodal retrieval tasks.

The model card notes that EmbeddingGemma 2 has not undergone post-training alignment, safety tuning, or output-level moderation, but its pre-training data was filtered for child sexual abuse material at multiple stages and automated filtering of personal information and other sensitive data. The training data spanned web documents, code, images, video, audio, and paired cross-modality samples in over 140 languages up to January 2025.

Google warns that while the model supports more than 100 languages, performance may not be equal across them, and omitting recommended task instruction prefixes on text inputs reduces embedding precision. The company also advises developers to adhere to the Gemma Prohibited Use Policy and implement application-level safeguards such as retrieval filtering and fairness testing.

On-device performance is a key aspect of EmbeddingGemma 2, with Google reporting that it requires around 191 megabytes of active RAM for text-only weights and about 567 megabytes for the full multimodal model. Weights are available on Hugging Face and Kaggle, with on-device-optimized versions through the LiteRT Community on Hugging Face. The model runs through various frameworks, including transformers, sentence-transformers 6.1.0 or later, MLX, vLLM, llama.cpp, SGLang, Ollama, and LMStudio.

Google AI Edge’s MediaPipe and LiteRT handle cross-platform deployment, with a MediaPipe Decision Task API supporting real-time classification and routing on multimodal context. Demos including Instant Media Search and Video Moments Finder ship in the Google AI Edge Gallery app, showcasing the capabilities of EmbeddingGemma 2. The model is also paired with Gemma 4 for contextual reasoning in the Google AI Edge Foresight app.

Availability in the Gemini Enterprise Agent Platform Model Garden is coming soon, according to Google. With its impressive performance and versatility, EmbeddingGemma 2 marks a significant milestone in multimodal embeddings and has far-reaching implications for data analysis tools and AI-generated image applications.