[{"data":1,"prerenderedAt":14},["ShallowReactive",2],{"artikel-en-berita-embeddinggemma-2":3},{"slug":4,"lang":5,"title":6,"description":7,"date":8,"updated":9,"topic":10,"product":9,"cover":11,"readMin":12,"html":13},"berita-embeddinggemma-2","en","AI news: Google releases EmbeddingGemma 2 for search on your own device","On 6 October 2026 Google launched EmbeddingGemma 2, a 740 million parameter open model that maps text, code, images, video and audio into one search space. It is not a chatbot.","2026-10-07",null,"berita","\u002Fimages\u002Fartikel\u002Fberita-embeddinggemma-2.webp",2,"\u003Cp>On 6 October 2026, Google DeepMind launched \u003Cstrong>EmbeddingGemma 2\u003C\u002Fstrong>. It is a small open model built for search, not for chatting the way ChatGPT does.\u003C\u002Fp>\n\u003Cp>You give it text, an image, audio, or video. The model turns that input into a list of numbers, called an embedding, so a computer can find items with a similar meaning even when the words do not match.\u003C\u002Fp>\n\u003Ch2>What Google announced\u003C\u002Fh2>\n\u003Cp>The official post is by Sahil Dua and Henrique Schechter Vera. Main facts from that post:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>The model has \u003Cstrong>740 million parameters\u003C\u002Fstrong>, is built on the Gemma 4 architecture, and is released under the \u003Cstrong>Apache 2.0\u003C\u002Fstrong> license. That license allows commercial use.\u003C\u002Fli>\n\u003Cli>It maps \u003Cstrong>text, code, images, video and audio\u003C\u002Fstrong> into one shared space. Google's examples: find a specific video clip from a voice memo, or search hours of audio with a text query.\u003C\u002Fli>\n\u003Cli>For text only, it can run at about \u003Cstrong>270 million\u003C\u002Fstrong> parameters. The vision encoder (170 million) and audio encoder (300 million) are optional.\u003C\u002Fli>\n\u003Cli>Output vectors start at \u003Cstrong>768\u003C\u002Fstrong> dimensions, and developers can truncate them to 512, 256, or 128. Google says this can cut storage by up to 6 times.\u003C\u002Fli>\n\u003Cli>The context window is \u003Cstrong>8K tokens\u003C\u002Fstrong>, four times the first EmbeddingGemma. Google says that covers about 5.5 minutes of audio, 29 images, or 58 video frames.\u003C\u002Fli>\n\u003Cli>Weights are on Hugging Face and Kaggle. Google said Model Garden availability is coming soon.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Google also said the original EmbeddingGemma has passed \u003Cstrong>20 million downloads\u003C\u002Fstrong>. That figure is Google's claim about the earlier model, not about EmbeddingGemma 2.\u003C\u002Fp>\n\u003Cp>On quality, Google reported its MTEB code score rising from 68.76 to 78.68. That is their own measurement. Independent tests still need to show whether the number holds outside Google's lab.\u003C\u002Fp>\n\u003Ch2>What this means for you\u003C\u002Fh2>\n\u003Cp>EmbeddingGemma 2 is not a box where you type a question and get an essay. It is the search layer behind an app.\u003C\u002Fp>\n\u003Cp>For everyday users, the near-term effect is phone or laptop apps that can search files, photos and recordings \u003Cstrong>without sending that data to a server\u003C\u002Fstrong>. Google points to examples in the Google AI Edge Gallery app: instant media search, and a finder for moments inside a video.\u003C\u002Fp>\n\u003Cp>You do not need to download this model today for daily work. If you build apps, or you want private search over your own files, this is the one to bookmark. If you only want to write or plan content, keep using the chat tool you already have. A shortlist is on the \u003Ca href=\"\u002Fen\u002Fai-tools\">AI tools\u003C\u002Fa> page.\u003C\u002Fp>\n\u003Ch2>Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Ftechnology\u002Fdevelopers-tools\u002Fembeddinggemma-2\u002F\" target=\"_blank\" rel=\"noopener\">EmbeddingGemma 2\u003C\u002Fa> (Google)\u003C\u002Fli>\n\u003C\u002Ful>\n",1793543302656]