Google 开源端侧多模态嵌入模型
推荐指数 76.0 NO. 010 · 2026.10.07
发布2026/10/06Score153Comments20
为什么值得看
EmbeddingGemma 2 是 Google 基于 Gemma 4 架构推出的 740M 参数多模态嵌入模型，支持文本、代码、图像、视频、音频统一嵌入。Apache 2.0 商用许可，专为消费级硬件上的本地搜索和隐私 RAG 设计，前代下载量已超 2000 万次。
编辑判断
端侧嵌入市场之前基本是 text-only 的天下，Multimodal 嵌入要么太大跑不动（如 OpenAI CLIP 系列），要么许可受限没法商用。EmbeddingGemma 2 的 740M 参数卡在了一个甜点区：比 BGE-M3 多了图像/视频能力，又比 Jina-CLIP 的 8B 版本轻量一个数量级。
做端侧 AI 搜索或隐私 RAG 的团队，之前处理图片/视频通常要调云端多模态 API 做预处理，现在可以全链路本地化了。不过要注意，统一嵌入空间的检索精度在跨模态场景下通常弱于专用模型，建议先在自己数据上跑召回率测试再替换现有管线。
最该试的人：正在用 Ollama + llama.cpp 做本地知识库、但卡在图片/视频检索这一步的开发者。
社区反馈
正面 20 条评论
核心争论：端侧多模态嵌入模型能否真正替代云端方案，以及参数分配与训练方法的权衡
Finally. I was getting annoyed that there's been an inflection point in how LLMs/agents work but there hasn't been a good moderate-size embeddings model, and this one is multimodal too! 270M for text only is great compared to older embedding models, and a total 440M for text + vision is also fa
Not just vision with video, but also audio, it really seems amazing. I’m not sure how it can handle vicinity of pairs of embeddings with for example some words and the audio where they’re spoken or an image where the text is handled. Building local multimodal search with this would be amazing. I’ve
> Building local multimodal search with this would be amazing. (lawyer here) — I’m curious: for what you’re describing, wouldn’t the machine need to be constantly running/updating the embeddings to take updated and new files into account? If so, how would that computation load compare to, say,