Embedding Gemma 2
A lightweight, open-source multimodal embedding model released by Google, designed to run offline on mobile devices. It supports retrieval across text, images, video, and audio, and is targeted at on-device semantic search scenarios where user privacy is a primary concern.
Core Facts
Google DeepMind has released Embedding Gemma 2, an open multimodal embedding model with 740M parameters that maps text, images, video, and audio into a shared vector space.
Google released Embedding Gemma 2, its first native multimodal open embedding model targeting on-device use, with only 740 million parameters, open weights under the Apache 2.0 license, capable of understanding text, code, images, audio, and video within a single shared space.
Timeline (last 90 days)
Embedding Gemma 2 has 740 million parameters, consisting of a 270M text encoder, a 170M vision encoder, and a 300M audio encoder, with runtime dynamic memory of approximately 191MB to 567MB; the 270M version requires a minimum of 0.5GB memory to run, and the full multimodal version can run locally with 1GB of memory.
Google DeepMind has released Embedding Gemma 2, an open multimodal embedding model with 740M parameters that maps text, images, video, and audio into a shared vector space.
Embedding Gemma 2 outputs 768-dimensional vectors (truncatable to 512, 256, or 128 dimensions), has a context length of 8,192 tokens, supports over 100 languages, and is released under the Apache 2.0 license.
Google open-sourced Embedding Gemma 2, which supports offline on-device retrieval of text, images, video, and audio, consuming approximately 567 MB of memory for multimodal retrieval in tests on the Pixel 11 Pro.
Google released Embedding Gemma 2, its first native multimodal open embedding model targeting on-device use, with only 740 million parameters, open weights under the Apache 2.0 license, capable of understanding text, code, images, audio, and video within a single shared space.
In a local test on an NVIDIA DGX Spark, Embedding Gemma 2 indexed 300 animal photos and hit 165 out of 200 retrieval queries, achieving an accuracy of 82.5%.
All Facts (6)
Google DeepMind has released Embedding Gemma 2, an open multimodal embedding model with 740M parameters that maps text, images, video, and audio into a shared vector space.
70%UnverifiedEmbedding Gemma 2 has 740 million parameters, consisting of a 270M text encoder, a 170M vision encoder, and a 300M audio encoder, with runtime dynamic memory of approximately 191MB to 567MB; the 270M version requires a minimum of 0.5GB memory to run, and the full multimodal version can run locally with 1GB of memory.
50%UnverifiedEmbedding Gemma 2 outputs 768-dimensional vectors (truncatable to 512, 256, or 128 dimensions), has a context length of 8,192 tokens, supports over 100 languages, and is released under the Apache 2.0 license.
50%UnverifiedIn a local test on an NVIDIA DGX Spark, Embedding Gemma 2 indexed 300 animal photos and hit 165 out of 200 retrieval queries, achieving an accuracy of 82.5%.
50%Partially VerifiedGoogle released Embedding Gemma 2, its first native multimodal open embedding model targeting on-device use, with only 740 million parameters, open weights under the Apache 2.0 license, capable of understanding text, code, images, audio, and video within a single shared space.
75%UnverifiedGoogle open-sourced Embedding Gemma 2, which supports offline on-device retrieval of text, images, video, and audio, consuming approximately 567 MB of memory for multimodal retrieval in tests on the Pixel 11 Pro.
50%