Google Open-Sources EmbeddingGemma 2, a Multimodal Embedding Model That Runs on Phones
谷歌发布开源多模态嵌入模型 EmbeddingGemma 2，参数量 740M，首次将 EmbeddingGemma 从纯文本扩展到图像、音频与视频，全部模态共享同一嵌入空间，体积小到可在手机端本地运行。模型已上架 Hugging Face。
已确认
- 模型为 740M 参数的轻量多模态嵌入模型，将文本、代码、图像、视频、音频映射到统一嵌入空间，上下文窗口 8K（比纯文本版更大）
- 采用 Apache 2.0 许可，可商用；支持 1 比特量化，仅损失 4.6% 检索质量，面向端侧场景
- 可 100% 免费在 Mac 上本地运行，无需云端调用（@SabooShubham 演示），适合本地语义检索、RAG 等零成本嵌入方案
- 基于今年 4 月发布的 Gemma 系列
- 此前第一代 EmbeddingGemma 带动了一批 AI for Science 项目落地，包括医学版 Embedding Gemma 3
为什么重要
- @ivanfioravanti 评价称 EmbeddingGemma 2 是嵌入模型界的"DeepSeek 时刻"，意指其以开源、高性价比方式冲击嵌入模型市场格局
- 对端侧开发者而言，免费本地跑多模态嵌入意味着检索、RAG 场景可摆脱云端 API 成本与延迟
2026-10-11 ~ 2026-10-12 · 6 related posts
- Episode 1: Unsloth Releases EmbeddingGemma-2 Multimodal Embedding Model(2026-10-06, 2 posts)
- Episode 2: Google Open-Sources EmbeddingGemma 2, a Native Multimodal On-Device Embedding Model(2026-10-06, 62 posts)
- Episode 3: Google EmbeddingGemma 2 Runs Locally in Browsers via WebGPU(2026-10-07, 3 posts)
- Episode 4: EmbeddingGemma 2 Delivers Unified Multimodal Search with 740M Parameters(2026-10-07, 3 posts)
- Episode 5: Google Open-Sources EmbeddingGemma 2, a 740M Multimodal Embedding Model(2026-10-09, 3 posts)
- Episode 6: Open-EmbeddingGemma Reproduces Google's 740M Embedding Model in PyTorch(2026-10-10, 3 posts)
- Episode 7: Google Open-Sources EmbeddingGemma 2, a Multimodal Embedding Model That Runs on Phones(2026-10-11, 6 posts)
Primary sources
- Google's EmbeddingGemma 2 runs free and fully local on a Mac — Saboo_Shubham_ ·
- Google's EmbeddingGemma 2 unifies text, image, video embeddings; 1-bit quantization keeps 95% of search quality — solyarisoftware ·
- Google ships EmbeddingGemma 2: 740M multimodal embeddings that run on phones — dl_weekly ·
- [source] Google's EmbeddingGemma 2 runs free and fully local on a Mac — Saboo_Shubham_ · 2026-10-11
- [source] Google ships EmbeddingGemma 2: 740M multimodal embeddings that run on phones — dl_weekly · 2026-10-11
- Google's Embedding Gemma 2 called the 'DeepSeek moment' for embedding models — HankYeomans · 2026-10-11
- [source] Google's EmbeddingGemma 2 unifies text, image, video embeddings; 1-bit quantization keeps 95% of search quality — solyarisoftware · 2026-10-12
- EmbeddingGemma 2 lands on Hugging Face, giving science projects a free accuracy boost — HildeKuehne · 2026-10-12
1 near-duplicate retellings: Saboo_Shubham_