Glonce

Google releases EmbeddingGemma 2, an open 740M multimodal embedding model

modelsThe Glonce editorsGoogle DeepMind🔥70Glonce rating70Impact60Novelty65Relevance100Surprise35Players85Credibility90
Google releases EmbeddingGemma 2, an open 740M multimodal embedding model

Google has launched EmbeddingGemma 2, a follow-up to the text-only EmbeddingGemma it introduced last year. The original model, aimed at helping apps organize, search and connect information on consumer hardware, passed 20 million downloads, and the announcement says builders used it for on-device search tools and privacy-first retrieval augmented generation (RAG) pipelines.

The new model moves beyond text and puts code, images, video and audio into a shared embedding space, handled by a single natively multimodal model. It is built on the Gemma 4 architecture, has 740 million parameters, and is released under a commercially permissive Apache 2.0 license. Google says it is built from the same technology as the Gemini Embedding models. Its pitch for the multimodal design: find a specific video clip from a voice memo, or search hours of audio recordings with a text query.

Google lists five properties. First, it claims leading scores among sub-1B multimodal embedders for its size on benchmarks such as MTEB Code and MAEB (Massive Audio Embedding Benchmark), while matching or outperforming many larger models across text, vision and audio tasks. Second, it is modular: text-only workloads need as little as 270M parameters, with an optional 170M vision encoder and a 300M audio encoder for full multimodal support. Third, with Matryoshka Representation Learning, developers can truncate output vectors from 768 dimensions down to 512, 256 or 128, which Google says gives up to 6x storage reduction for local vector databases and memory usage. Fourth, with quantization on a Google Pixel 11 Pro, the model needs as little as about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model. Fifth, the context window is 8K tokens, 4x larger than EmbeddingGemma 1, enough for up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations.

On quality, Google says EmbeddingGemma 2 matches the multilingual text performance of EmbeddingGemma and improves MTEB Code by 9.92 points, from 68.76 to 78.68. It positions the model for local codebase indexing, semantic code search and coding agent retrieval. Across image, video, documents and audio, the post says it sets a new standard in quality-per-parameter for sub-1B models and outperforms some specialist models more than twice its size. Full evaluation metrics are in the model card.

The post stresses running everything locally: generating embeddings on the device helps ensure data privacy, reduces pipeline latency and allows cross-modal search and retrieval that works entirely offline. Paired with a generative model such as Gemma 4, the embedder can power on-device RAG over multimodal data. Because it shares its text tokenizer and audio encoder with Gemma 4, both models can run in one pipeline with a lower combined memory footprint.

For getting started, the weights are on Hugging Face and Kaggle, with Gemini Enterprise Agent Platform Model Garden availability coming soon, and on-device optimized models sit in LiteRT Community on Hugging Face. Google names Google AI Edge MediaPipe and LiteRT for app deployment, transformers.js and WebGPU for the browser, and transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio for serving. Qdrant is named for storing vectors, and Unsloth for fine-tuning guidance.

Key facts

  • EmbeddingGemma 2 has 740 million parameters, is built on the Gemma 4 architecture and is released under Apache 2.0, with text, code, images, video and audio in one shared embedding space.
  • It is modular: 270M parameters for text-only workloads, plus an optional 170M vision encoder and a 300M audio encoder.
  • Google reports a 9.92-point gain on MTEB Code over EmbeddingGemma, from 68.76 to 78.68, while matching its multilingual text performance.
  • With quantization on a Google Pixel 11 Pro, it needs about 191MB of active RAM for text-only weights and about 567MB for the full multimodal model.
  • The context window is 8K tokens (4x EmbeddingGemma 1), and vectors can be truncated from 768 to 512, 256 or 128 dimensions for up to 6x storage reduction.

Why it matters

The original EmbeddingGemma covered text only. This release puts code, images, video and audio into the same embedding space in one open model small enough, by Google's account, to run on a phone. That makes cross-modal local search possible, such as finding a video clip from a voice memo or searching audio with a text query, without sending data to a server. The code gain is concrete: MTEB Code rises from 68.76 to 78.68. The Apache 2.0 license and the Gemma 4 foundation also mean the model can sit alongside a generative model in the same on-device pipeline.

Who it affects

Developers building on-device search and RAG features for apps, browsers and edge hardware are the main audience, especially where privacy and offline operation matter. Teams working on local codebase indexing, semantic code search or coding agent retrieval are named as a fit. Builders who already use EmbeddingGemma for text can look at the multimodal upgrade, and anyone running a local vector database may care about the dimension truncation options.

How to use it

Download the weights from Hugging Face or Kaggle; a Gemini Enterprise Agent Platform Model Garden listing is coming soon. On-device optimized models are in LiteRT Community on Hugging Face. For apps, Google points to Google AI Edge MediaPipe for turnkey embedding and retrieval tasks, LiteRT for custom integration, and transformers.js or WebGPU for the browser. Serving works with transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio. Qdrant can store the vectors, and Unsloth provides fine-tuning guidance. To save space, truncate output vectors from 768 to 512, 256 or 128 dimensions with Matryoshka Representation Learning. The license is Apache 2.0.

How solid is it

This is a first-party announcement from Google, and the quality claims are its own. The only numeric benchmark result in the post is the MTEB Code improvement from 68.76 to 78.68. The 'leading scores among sub-1B multimodal embedders' claim for MAEB and the claim of beating specialist models more than twice its size come without numbers in the text, and the competing models are not named. The post points to the model card for full evaluation metrics. The memory figures are tied to a specific device, the Google Pixel 11 Pro, and to quantization.

Risks and caveats

The 6x storage reduction is an 'up to' figure, and the post does not say which dimension size it corresponds to or whether quality drops when vectors are truncated. The RAM numbers depend on quantization, and the post does not specify the scheme. No latency, throughput or energy figures are given. The 8K context caps a single input at roughly 5.5 minutes of audio, 29 images or 58 video frames. The Model Garden listing is not yet live. The post gives no pricing or hosted-API availability, only weight downloads.

“Generating embeddings locally helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline.”

— Google, EmbeddingGemma 2 announcement