Key Info

Google DeepMind announced EmbeddingGemma 2, a new open embeddings model designed for on-device use cases spanning code, image, video, audio, and text.

Highlights

  • Modular footprint ranges from 270M for text/code up to 740M for full multimodal
  • Supports Matryoshka embeddings, truncating down to 128 dimensions for 6x storage savings
  • Claims a +14% boost on MTEB Code
  • Hugging Face demos built with Gradio are available