Reldex
Back to releases
EmbeddingGemmaNew product

EmbeddingGemma 2 brings multimodal embeddings to local devices

Google DeepMind released EmbeddingGemma 2, an open embedding model for text and code, images, video, and audio. Its modular architecture has about 740M parameters and supports a shared 8,192-token input budget.

Announced ReleasedCheck sources ↓

The update, in brief

  • Outputs 768-dimensional vectors, with 512, 256 and 128-dimensional truncation; 128 dimensions are mainly suited to text-only tasks.
  • Use the 270M text component alone, or load the vision and audio encoders for multimodal retrieval.
  • Google reports about 191MB active RAM for quantized text weights and 567MB for the full model on Pixel 11 Pro; these are publisher figures, not our measurements.

Availability details

Public weights are available on Hugging Face and Kaggle under Apache 2.0; the model card also requires adherence to the Gemma Prohibited Use Policy. Model Garden availability is forthcoming. Media modalities share the same input budget.

Check sources

2

  1. Sahil Dua & Henrique Schechter Vera · Google DeepMindOfficial · ArticleEmbeddingGemma 2: an open, lightweight multimodal embedding modelblog.google
  2. Google DeepMindOfficial · Release notesEmbeddingGemma 2 model cardai.google.dev

More about this release

Demos, launch pages and discussions explicitly connected to this announcement.

Prepared by Reldex · Based on linked sources