Google DeepMind Google DeepMind 官方 Official 社区 showcase Community showcase

EmbeddingGemma 2：把文字、图片、声音和视频收成同一套向量的小模型 EmbeddingGemma 2: a small model that turns text, images, audio, and video into one vector space

EmbeddingGemma 2: a small model that turns text, images, audio, and video into one vector space

EmbeddingGemma 2 是 Google DeepMind 在 2026-10-06 发布的开源嵌入模型。嵌入是一串数字，用来表示一段文字、一张图或一段声音的意思，方便按意思搜索，而不是按关键词硬对。这一版把文字（含代码）、图片、视频和声音放进同一个 768 维的空间。模型卡写总参数 7.4 亿，其中文字部分约 2.7 亿，视觉编码器 1.7 亿，音频编码器 3 亿，可以只加载用得到的那一部分。许可页标题是 Apache License 2.0。它面向手机和笔记本上的本地搜索。模型卡里的测试分是 Google 自己报的，本库没有复跑，也没有下载权重。 EmbeddingGemma 2 is an open embedding model Google DeepMind released on 2026-10-06. An embedding is a list of numbers that stands for the meaning of a passage, a picture, or a sound, so you can search by meaning rather than by exact keywords. This version puts text (including code), images, video, and audio in one 768-dimension space. The model card says 740 million parameters in total: about 270 million for text, a 170 million vision encoder, and a 300 million audio encoder, and you can load only the part you need. The license page is titled Apache License 2.0. It is aimed at search on a phone or laptop. Scores on the model card are Google’s own. This library did not re-run them and did not download the weights.

预览：Chrome 实拍 Google 官博首屏 Preview: Chrome capture of the Google blog above the fold

来源Sources

官方模型页 Official model page Google 官博：EmbeddingGemma 2 Google blog: EmbeddingGemma 2 Sahil Dua, Henrique Schechter Vera, Google DeepMind 2026年10月6日 Oct 6, 2026 https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/ 官方模型页 Official model page Hugging Face 模型卡：google/embeddinggemma-2 Hugging Face model card: google/embeddinggemma-2 Google DeepMind 2026年10月6日 Oct 6, 2026 https://huggingface.co/google/embeddinggemma-2 官方模型页 Official model page Gemma 文档：EmbeddingGemma Gemma docs: EmbeddingGemma Google AI for Developers https://ai.google.dev/gemma/docs/embeddinggemma 官方模型页 Official model page 许可页：Apache License 2.0 License page: Apache License 2.0 Google https://ai.google.dev/gemma/apache_2 社区 showcase Community showcase Hacker News 讨论 Hacker News thread Hacker News 2026年10月6日 Oct 6, 2026 https://news.ycombinator.com/item?id=49980487

为什么推荐Why we recommend it 规格和表内分数以模型卡为准。官博另有 Pixel 机型内存、5.5 分钟音频、29 张图、58 帧视频等说法，未写入要点，避免和模型卡的“数分钟音视频”并成两套数字。本库未下载权重，未做检索实验，未复跑 MTEB。同板没有另一张嵌入模型卡，因此 relatedModelIds 留空。 Specs and in-table scores follow the model card. The blog also states Pixel memory, 5.5 minutes of audio, 29 images, and 58 video frames. Those are not copied into the highlights, so they do not sit beside the model card’s “minutes of audio or video” as a second set of numbers. This library did not download weights, did not run a search experiment, and did not re-run MTEB. There is no other embedding card on this board, so relatedModelIds is empty.

核验记录：来源、抓取与截图说明Verification log: sources, checks, and captures

收录理由：HN item 49980487，标题 EmbeddingGemma 2: An open, lightweight multimodal embedding model，作者 ilreb，2026-10-06T16:03:49.000Z（约 2026-10-07 00:03 CST），撰写时 Algolia 352pts。官博 https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/ 撰写时 HTTP 200，页内日期 Oct 06, 2026，作者 Sahil Dua 与 Henrique Schechter Vera。模型卡 https://huggingface.co/google/embeddinggemma-2 撰写时 HTTP 200，front matter 写 license apache-2.0。卡内许可链接 https://ai.google.dev/gemma/docs/gemma_4_license 撰写时 301 到 https://ai.google.dev/gemma/apache_2 ，该页 HTTP 200，标题 Apache License 2.0。文档 https://ai.google.dev/gemma/docs/embeddinggemma HTTP 200。预览为 Chrome 实拍官博首屏裁 3:2 WebP。不打星。Why: HN item 49980487, title EmbeddingGemma 2: An open, lightweight multimodal embedding model, author ilreb, 2026-10-06T16:03:49.000Z (about 00:03 CST on 2026-10-07), 352pts on Algolia at write time. The blog https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/ returned HTTP 200, dated Oct 06, 2026, by Sahil Dua and Henrique Schechter Vera. The model card https://huggingface.co/google/embeddinggemma-2 returned HTTP 200, with license apache-2.0 in the front matter. The card’s license link https://ai.google.dev/gemma/docs/gemma_4_license returned 301 to https://ai.google.dev/gemma/apache_2, which returned HTTP 200 and is titled Apache License 2.0. Docs https://ai.google.dev/gemma/docs/embeddinggemma returned HTTP 200. Preview is a Chrome capture of the blog above the fold, cropped to 3:2 WebP. Unrated.

要点（官方称 / 社区称）Highlights (vendor / community claims)

模型卡写：总参数 7.4 亿。文字骨干 1.3 亿加嵌入层 1.4 亿，合起来约 2.7 亿；视觉编码器 1.7 亿，音频编码器 3 亿。输出默认 768 维，也可以截成 512、256 或 128 维，用来少占存储。上下文窗口 8192 token。输入可以是文字、图片、视频、声音，以及它们的组合。 The model card says 740 million parameters total. The text side is a 130 million backbone plus a 140 million embedder, about 270 million together. The vision encoder is 170 million and the audio encoder is 300 million. The default output is 768 dimensions, and it can be cut to 512, 256, or 128 to store less. The context window is 8,192 tokens. Inputs can be text, images, video, audio, and combinations of those.
模型卡上的对比分（官方数字，本库未复跑）：代码检索 MTEB Code 的 NDCG@10，EmbeddingGemma 2 是 78.68，上一版是 68.76。多语言文本 MTEB 的平均分是 61.36，上一版是 61.15。模型卡正文另写代码任务相对上一版大约提高 14%。这两处是同一张表和同一页的说法。 Comparison scores on the model card (vendor numbers, not re-run here): on MTEB Code, NDCG@10 is 78.68 for EmbeddingGemma 2 and 68.76 for the previous version. Multilingual text MTEB mean is 61.36 versus 61.15. The card’s prose also says code tasks improved by about 14% versus the previous version. Those are the table and the prose on the same page.
怎么拿（官方渠道，本库没下载）：权重在 Hugging Face 的 google/embeddinggemma-2。文档在 ai.google.dev 的 EmbeddingGemma 页。模型卡写可以用 transformers、sentence-transformers 等常见库加载，也面向设备端。官博还写了在手机上的内存占用和能塞进多少分钟音频，那些设备数字没有写进模型卡的规格表，本卡不另抄一份。 Where to get it (official channels, not downloaded here): weights are at google/embeddinggemma-2 on Hugging Face. Docs are the EmbeddingGemma page on ai.google.dev. The model card says it loads with common libraries such as transformers and sentence-transformers, and that it is meant for devices. The blog also states phone memory and how many minutes of audio fit. Those device figures are not in the model card’s spec table, so this card does not copy a second set.