Google Open-Sources EmbeddingGemma 2, Its First Natively Multimodal Embedding Model

Google CEO Sundar Pichai and Google DeepMind officially released EmbeddingGemma 2 around October 7. It is Google's first natively multimodal open-source embedding model and one of the few lightweight multimodal embedding solutions designed for on-device deployment, making it worth attention.

Confirmed

• Official positioning: EmbeddingGemma 2 is DeepMind's first natively multimodal, open-source embedding model designed for on-device efficiency, going beyond text-only embeddings.
• Scale: With only 740M parameters, it uses a modular encoder to map text (including code), images, video, and audio into a single embedding space (768 dimensions), and supports Matryoshka Representation Learning.
• Openness: Open-sourced under the Apache 2.0 license, with weights available on Hugging Face and Kaggle; unsloth has already provided a GGUF version.
• Use cases: Primarily aimed at offline, privacy-first RAG for on-device deployment.

Why it matters

• Multimodal embedding models have previously relied on large cloud-based models or separate single-modality models. EmbeddingGemma 2 unifies vector representations across five modalities in a single 740M model, dramatically lowering the barrier to on-device retrieval and RAG deployment.
• The Apache 2.0 license, combined with quick availability on Hugging Face/Kaggle and community GGUF versions, means developers can try it on local devices right away, accelerating privacy-first offline AI applications.

2026-10-06 ~ 2026-10-07 · 33 related posts

• Google releases EmbeddingGemma 2: 740M multimodal embeddings for on-device RAG — jacek2023 · 2026-10-06
• [source] Google DeepMind details EmbeddingGemma 2: unified embeddings for code, image, audio, video — GoogleDeepMind · 2026-10-07
• [source] EmbeddingGemma 2 Maps Text, Image, Video & Audio Into One Shared 768-Dim Space, 740M Params — tomaarsen · 2026-10-07
• One text query retrieves photos, audio and video in EmbeddingGemma 2's shared space — tomaarsen · 2026-10-07
• EmbeddingGemma 2 is modular: load only 270M for text-only on phones and laptops — tomaarsen · 2026-10-07
• EmbeddingGemma 2 code retrieval jumps 14%, multilingual gains marginal over v1 — tomaarsen · 2026-10-07
• EmbeddingGemma 2 multimodal scores: 67.84 visual docs, 50.67 video, 69.54 audio — tomaarsen · 2026-10-07
• EmbeddingGemma 2 dim tradeoff: 128d shrinks vectors 6x but MMEB drops to 45.65 — tomaarsen · 2026-10-07
• Matryoshka training lets embeddings truncate to 256d with a third of the storage — tomaarsen · 2026-10-07
• Task prefixes for embedding: pair SearchQuery with Document for retrieval — tomaarsen · 2026-10-07
• One vector for a whole product listing: multimodal embedding mixes text, photos and video — tomaarsen · 2026-10-07
• Google's multimodal embedding: images cost 280 tokens, audio 25 tokens per second — tomaarsen · 2026-10-07
• Don't Run This Model in FP16: Activations Exceed Its Dynamic Range — tomaarsen · 2026-10-07
• Never run EmbeddingGemma 2 in FP16: use BF16 or FP32 to avoid NaNs — tomaarsen · 2026-10-07
• Google ships Embedding Gemma 2 on Hugging Face — victormustar · 2026-10-07
• EmbeddingGemma 2 Packs 740M Params Into a Unified Embedding Space, Runs in Browser — nicodotdev · 2026-10-07
• Google's multimodal embeddinggemma-2 trends on Hugging Face — google · 2026-10-07
• EmbeddingGemma 2 weights land on Hugging Face with multimodal feature extraction — minchoi · 2026-10-07
• Google open-sources EmbeddingGemma 2: 740M multimodal embeddings that run fully on-device — minchoi · 2026-10-07

14 near-duplicate retellings: sundarpichai · GoogleDeepMind · GlennCameronjr · osanseviero · osanseviero · gnukeith · tomaarsen · tomaarsen · DynamicWebPaige · Recoil42 · rseroter · testingcatalog · minchoi · NVIDIAAI