Google 发布 EmbeddingGemma 2 论文，744M 参数端侧多模态嵌入模型
Google released its EmbeddingGemma 2 paper.
Google 发布 EmbeddingGemma 2 论文，这是一个 Apache 2.0 许可的端侧多模态嵌入模型，把文本、代码、图像、音频和视频映射到同一向量空间。模型基于 Gemma 4 构建，通过向更大的 Gemini Embedding 2 学习并用对比训练调优，744M 参数规模在多项任务上超过大 2 至 12 倍的模型。在 Pixel 11 Pro 上，量化后的文本权重占用约 191MB 活跃内存，完整多模态模型约 567MB，上下文窗口扩大 4 倍至 8K tokens，可容纳约 5.5 分钟音频、29 张图像或 58 帧视频；直接嵌入语音在 MSEB 口语查询检索基准上得分 69.54，高于先转写再检索的 57.59，代码检索在 MTEB Code 上从 EmbeddingGemma 1 的 68.76 提升到 78.68。
推荐理由
AI 评分 62
744M 参数的开源嵌入模型把文本、图像、音频、视频放进同一向量空间，并给出手机端内存占用与检索基准对比。
主题
正文
Google released its EmbeddingGemma 2 paper.
embedding model for on-device multimodal AI, under an Apache 2.0 license.
> puts text, code, images, audio and video into 1 searchable space on phones. gives phones a missing piece: a way to understand and search your own stuff without sending it to a server.
> shows that a 744M-parameter open model can embed text, images, audio, and video, beating many models 2 to 12 times larger.
> On a Pixel 11 Pro, quantized text weights take about 191MB of active RAM, and the full multimodal model takes about 567MB.
> The context window grows 4x to 8K tokens, enough for roughly 5.5 minutes of audio, 29 images or 58 video frames in 1 input.
They built EmbeddingGemma 2 on Gemma 4, trained it to copy the larger Gemini Embedding 2, then tuned it with contrastive training. All media types share 1 vector space, and the full model fits in 567 MB on a Pixel 11 Pro with 4-bit weights.
Embedding speech directly beat transcribing it, 69.54 versus 57.59 on the MSEB spoken-query retrieval benchmark. Code search scored 78.68 on MTEB Code, up from 68.76 for EmbeddingGemma 1.