{"id":38419,"date":"2025-06-27T21:18:58","date_gmt":"2025-06-27T13:18:58","guid":{"rendered":"https:\/\/www.1ai.net\/?p=38419"},"modified":"2025-06-27T21:18:58","modified_gmt":"2025-06-27T13:18:58","slug":"%e8%b0%b7%e6%ad%8c%e6%ad%a3%e5%bc%8f%e5%8f%91%e5%b8%83-gemma-3n-%e5%b0%8f%e9%92%a2%e7%82%ae%e6%a8%a1%e5%9e%8b%ef%bc%9a2gb-%e5%86%85%e5%ad%98%e6%9c%ac%e5%9c%b0%e7%8e%a9%e8%bd%ac-ai-%e5%a4%9a%e6%a8%a1","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/38419.html","title":{"rendered":"Google officially releases Gemma 3n mini-steel model: 2GB of RAM to play AI multimodal locally"},"content":{"rendered":"<p>June 27 - Technology outlet NeoWin published a blog post today (June 27) reporting that after previewing the release at the 2025 I\/O developer conference, the<a href=\"https:\/\/www.1ai.net\/en\/tag\/%e8%b0%b7%e6%ad%8c\" title=\"[View articles tagged with [Google]]\" target=\"_blank\" >Google<\/a>Officially launched <a href=\"https:\/\/www.1ai.net\/en\/tag\/gemma-3n\" title=\"[See articles with [Gemma 3n] labels]\" target=\"_blank\" >Gemma 3n<\/a> <a href=\"https:\/\/www.1ai.net\/en\/tag\/%e7%ab%af%e4%be%a7%e5%a4%9a%e6%a8%a1%e6%80%81%e6%a8%a1%e5%9e%8b\" title=\"[Sees articles containing [end-side MIM] labels]\" target=\"_blank\" >End-side multimodal modeling<\/a>,<strong>Supports running locally on cell phones, tablets and laptops and handles multiple data types of audio, text, images and video.<\/strong><\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-38420\" title=\"f09f02bcj00syinm20082d000v900hkp\" src=\"https:\/\/www.1ai.net\/wp-content\/uploads\/2025\/06\/f09f02bcj00syinm20082d000v900hkp.jpg\" alt=\"f09f02bcj00syinm20082d000v900hkp\" width=\"1125\" height=\"632\" \/><\/p>\n<p>Compared to the preview released in May, the latest full release of Gemma 3n further improves performance by supporting local operation on hardware with 2GB of RAM, with a focus on coding and inference.<\/p>\n<p>1AI cites a blog post that describes Gemma 3n as having two scaled versions, E2B with 5 billion (5B) parameters supporting operation on devices with more than 2GB of memory, and E4B with 8 billion (8B) parameters supporting operation on devices with more than 3GB of memory, both of which, through architectural innovations, have memory footprints equivalent to the 2 billion (2B) and 4 billion (4B) models.<\/p>\n<p>In terms of architecture, Gemma 3n innovates with the MatFormer architecture to provide computational flexibility, in addition to the use of Per Layer Embeddings (PLE) for memory efficiency, the MobileNet-v5 visual coder, and more.<\/p>\n<p>For the MatFormer architecture, Google describes it with a Russian nesting doll analogy: a larger model contains a smaller but fully functional version inside, allowing a single model to run at different sizes depending on the task.<\/p>\n<p>Gemma 3n achieves quality improvements in multilingualism (text support in 140 languages and multimodal understanding in 35 languages), math, coding, and reasoning.<\/p>\n<p>In terms of performance benchmarks, the larger E4B model is the first model with less than 10B parameters but an LMArena score of over 1300.<\/p>\n<p>The model's audio capabilities now support speech-to-text and translation on the device, using an encoder capable of handling detailed speech.<\/p>\n<p>The visual side is powered by a new encoder called MobileNet-V5, which is faster and more efficient than its predecessor. It can process video at up to 60FPS on Google Pixel devices.<\/p>","protected":false},"excerpt":{"rendered":"<p>On June 27, according to the news, NeoWin, a technology medium, published today (27 June), reports that, following a preview at the I\/O Developer ' s Congress in 2025, Google has officially launched the Gemma 3n end multi-modular model, which supports local running on mobile phones, tablets and laptops, processing multiple data types of audio, text, pictures and videos. Compared to the preview release in May, the latest release of the Gemma 3n full version further enhanced performance, supported local running on the hardware of the 2GB memory, and focused on enhancing coding and reasoning capabilities. 1AI quotes Boven, Gemma 3n with two size versions, E2B with 5 billion (5B) parameters<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[146],"tags":[6708,6707,281],"collection":[],"class_list":["post-38419","post","type-post","status-publish","format-standard","hentry","category-news","tag-gemma-3n","tag-6707","tag-281"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/38419","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=38419"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/38419\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=38419"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=38419"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=38419"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=38419"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}