Google DeepMind Launches EmbeddingGemma 2 for On-Device AI

Google DeepMind on Oct. 6, 2026, released EmbeddingGemma 2, a multimodal embedding model that can compare a voice memo to a video clip, search a codebase with plain English, or retrieve images using an audio query — all without sending a single byte to the cloud. Built on the Gemma 4 architecture and licensed under the commercially permissive Apache 2.0 license, it marks a significant expansion over its text-only predecessor.

The original EmbeddingGemma, a 308-million-parameter text-only model, racked up more than 20 million downloads after its launch. According to Google DeepMind, the developer response “blew past” the team’s expectations, with builders using it to power on-device search tools and privacy-first retrieval augmented generation (RAG) pipelines. Version 2 keeps the same zero-cost, offline-capable approach while dramatically expanding scope.

What’s Actually New

EmbeddingGemma 2 has 740 million total parameters, but its modular design means it can run text-only workloads using as little as 270M parameters, with optional vision (170M) and audio (300M) encoders added as needed. All four modalities — text, code, images, audio and video — map into a single shared 768-dimensional embedding space, which means a text query and an audio clip can be compared directly using cosine similarity without any bridging model in between.

The context window grew from 2K tokens in EmbeddingGemma 1 to 8K tokens — a 4x increase — which translates to processing up to 5.5 minutes of audio, 29 images or 58 video frames in a single pass on local hardware. With quantization on a Google Pixel 11 Pro, the full multimodal model requires roughly 567MB of active RAM; the text-only configuration drops to about 191MB.

Code search saw the sharpest benchmark jump: a 9.92-point improvement on the MTEB Code benchmark (from 68.76 to 78.68), roughly a 14% gain over version 1. Multilingual text performance stayed essentially flat at 61.36 versus 61.15, reflecting a deliberate focus on new modalities rather than re-engineering what already worked. The model also supports more than 100 languages out of the box.

Using a technique called Matryoshka Representation Learning, developers can shrink output vectors from 768 dimensions down to 512, 256 or 128 without retraining, yielding up to a 6x reduction in local vector database storage — a meaningful advantage when running everything on a laptop or phone.

How It Stacks Up Against the Competition

The multimodal embedding market is dominated by cloud APIs with per-token or per-image pricing: Google’s own Gemini Embedding 2 charges around $0.20 per million text tokens; Voyage Multimodal 3.5 from Voyage AI charges roughly $0.60 per billion pixels plus token costs; and Cohere Embed v4 runs about $0.47 per million image tokens. None of these run locally.

The most direct open-weight alternative at a comparable scale is Jina Embeddings v5 Omni from Jina AI, which supports text, image, video and audio — but carries a CC BY-NC license that bars commercial use. EmbeddingGemma 2’s Apache 2.0 license is a real differentiator for anyone building a product they intend to ship or sell.

On raw benchmark scores at larger scales, EmbeddingGemma 2 faces stiffer competition. Models like WeMM-Embedding-2B achieve an MMEB-v2 score of 77.9, and a 9B variant tops the official leaderboard at 80.6. EmbeddingGemma 2 scores 57.28 on MMEB v2 image retrieval, so larger open models do outperform it in that dimension. Google DeepMind’s explicit design constraint, however, is the sub-1B, on-device tier, where the company says it achieves leading scores on MTEB Code and the Massive Audio Embedding Benchmark while matching or beating several larger models across text, vision and audio tasks.

Why This Matters for Students and Early-Career Developers

Until now, building a search tool that could unify text, images, audio and video typically meant stitching together multiple API calls — each billed separately — and routing user data through third-party servers. EmbeddingGemma 2 collapses that into a single model that runs entirely offline.

“Generating embeddings locally helps ensure data privacy, reduces pipeline latency, and empowers developers to build cross-modal search and retrieval that works entirely offline,” wrote Sahil Dua and Henrique Schechter Vera, research engineers at Google DeepMind.

A student can now build a personal media search engine, an offline code-search tool for their own GitHub repositories, or a voice-memo-to-video-clip finder — all as an on-device RAG pipeline paired with Gemma 4, which shares the same text tokenizer and audio encoder for a lower combined memory footprint — entirely on a laptop, with no API key and no monthly bill. The improved code embedding score makes semantic code search a viable portfolio project using only free, open-source tooling.

The model is available immediately on Hugging Face and Kaggle and works with popular developer tools including sentence-transformers, Ollama, llama.cpp, vLLM and LMStudio. For international students building for non-English-speaking communities, broad multilingual support is included without any extra configuration.

For job seekers looking to stand out, a local multimodal search project built with EmbeddingGemma 2 demonstrates real applied AI skills — RAG pipelines, vector databases, on-device inference — without requiring cloud credits or a company budget. That’s a meaningful entry point into a field where hands-on projects increasingly matter as much as coursework.

Source: Google DeepMind