{
  "id": 12468302,
  "title": "EmbeddingGemma 2",
  "url": "https://urgent.news/2026/10/06/embeddinggemma-2",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-06T16:03:49.000Z",
  "source": {
    "name": "Hacker News",
    "slug": "hacker-news",
    "url": "https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/"
  },
  "original_language": "en",
  "account": "Yesterday, we unveiled EmbeddingGemma 2, a versatile model that extends text embeddings to code, images, video, and audio. Developed on the Gemma 4 architecture and released under the Apache 2.0 license, EmbeddingGemma 2 boasts 740 million parameters, making it ideal for on-device use. This model can locate specific video clips from voice memos, search audio recordings using text queries, and more, all processed by a single, multimodal model.\n\nEmbeddingGemma 2 maintains the high-quality text performance of its predecessor while significantly improving code performance, achieving a 9.92-point increase in MTEB Code (from 68.76 to 78.68). This makes it an excellent choice for indexing local codebases, performing semantic code searches, and conducting coding agent retrievals.\n\nIn terms of performance, EmbeddingGemma 2 excels across various data types, setting a new standard for quality-per-parameter in sub-1B models and even surpassing some specialist models more than twice its size. For detailed evaluation metrics and model information, refer to the EmbeddingGemma 2 model card.\n\nBy generating embeddings locally, EmbeddingGemma 2 enhances data privacy, reduces pipeline latency, and enables developers to create cross-modal search and retrieval systems that operate entirely offline. When integrated with generative models such as Gemma 4, EmbeddingGemma 2 supports on-device Retrieval Augmented Generation (RAG) pipelines capable of understanding complex multimodal data.\n\nThe model's compatibility with the Gemma 4 architecture allows it to share the same text tokenizer and audio encoder, reducing the combined memory footprint when used alongside other models. To explore how to build on-device search and RAG systems with LiteRT, consult the Google AI Edge blog post.\n\nEmbeddingGemma 2 has been developed in partnership with several developers to ensure seamless integration into your existing workflows. For a comprehensive guide on building on-device search and RAG systems using LiteRT, check out the developer guide, documentation, and other resources provided.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Simon Willison",
        "title": "EmbeddingGemma 2",
        "url": "https://urgent.news/2026/10/06/embeddinggemma-2-12468416",
        "published": "2026-10-06T20:37:53.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}