zhineng.dev
English
Get started
← All models

EmbeddingGemma 2

Published Publisher Google

Overview

EmbeddingGemma 2 is an open-weight, lightweight multimodal embedding model from Google DeepMind, built on the Gemma 4 architecture and released under the Apache 2.0 license. It is designed for local and on-device inference rather than text generation. The full model has 740 million parameters; its modular design allows applications to load only the encoders they need. (blog.google)

Capabilities and uses

  • Maps text, code, images, audio, video and combinations of these inputs into a shared 768-dimensional embedding space, enabling cross-modal retrieval. (ai.google.dev)
  • Supports more than 100 languages and adjustable embedding dimensions of 128, 256, 512 or 768 using Matryoshka Representation Learning. Smaller vectors reduce storage but can affect retrieval quality. (ai.google.dev)
  • Suitable for semantic search, multimodal retrieval-augmented generation, classification, clustering and code search. It can run locally and offline on supported consumer hardware. (ai.google.dev)
  • The text/code model has 270 million parameters; the optional vision and audio encoders add 170 million and 300 million respectively. (ai.google.dev)

Limits and considerations

  • All modalities share an 8,192-token context window. At default settings, a single-modality input can contain approximately 29 images, 58 video frames or 327 seconds of audio; mixed inputs share the same budget. Video is sampled at one frame per second by default, and audio should be 16 kHz mono. (ai.google.dev)
  • Performance can vary across languages and domains, and embeddings can reflect biases in training data. The model is not safety-tuned or moderated at the output level; deployers should evaluate downstream retrieval quality, fairness and safeguards. (ai.google.dev)
  • Google's reported quantized memory figures—approximately 191 MB active RAM for text-only weights and 567 MB for the full model—are specific to a Pixel 11 Pro and should not be assumed for every device. (blog.google)

Pricing

As of November 15, 2026, a Google-hosted API price for this model could not be verified: currency, input-token charges and output charges are unverified. Open-weight local deployment is available under Apache 2.0, but users may incur their own hardware, electricity or hosting costs. No third-party hosting price is presented as Google's price. (ai.google.dev)

Release and official sources