Google Released EmbeddingGemma 2 Multimodal Model

The lightweight architecture extends Google's local-first embedding capabilities for resource-constrained hardware.

Updated on Oct. 6, 2026 in Artificial Intelligence

Bold flat-color editorial illustration of a silicon wafer and a geometric crystal, representing high-precision local data processing. — Google released EmbeddingGemma 2, a lightweight multimodal model designed to perform data vectorization directly on local edge hardware. AI Illustration. Upload story photo >

Live Poll

Do you trust on-device AI models to keep your personal data more secure?

Google has launched EmbeddingGemma 2, a new iteration of its lightweight multimodal model. The update builds upon the initial 308-million parameter model unveiled by Google DeepMind on September 4, 2025.

Why it matters

The model is designed to facilitate privacy-sensitive applications by generating embeddings directly on local hardware. It addresses the need for efficient, low-memory content vectorization in edge computing environments.

The original EmbeddingGemma featured a 2K token context window and supported over 100 languages with sub-200MB RAM usage. It achieved inference times between 15 and 22 milliseconds on EdgeTPU hardware using output dimensions ranging from 128 to 768.

The players

Google

A global technology company focused on search, cloud computing, and AI research through its DeepMind division.

Google DeepMind

A research laboratory that develops the Gemma family of open models and advanced neural network architectures.

The details

EmbeddingGemma models function by converting diverse content types into numerical fingerprints, known as embeddings, which represent data in a high-dimensional space. The original model utilized quantization-aware training—a process that reduces the precision of model weights to enable lower memory footprints without sacrificing significant accuracy. This allows the system to perform complex similarity searches or classification tasks locally on devices without needing cloud connectivity.

September 4, 2025: Google DeepMind unveiled the first EmbeddingGemma model.
September 24-25, 2025: A paper on the first EmbeddingGemma model was released.
October 6, 2026: Information regarding the launch of EmbeddingGemma 2 was published.

The Tech Race

This release follows the rapid development of the Google Gemma model family as a standard for high-performance, local-run AI. It marks a continued effort to compete with proprietary cloud-based embedding services by offering comparable capabilities on standard mobile or edge hardware.

Developers working on privacy-focused applications will gain access to tools that perform content vectorization without sending user data to the cloud. The model is specifically optimized for hardware like EdgeTPU, meaning it will likely appear in mobile and IoT devices first.

The takeaway

EmbeddingGemma 2 underscores the industry-wide shift toward offloading complex AI inference from massive data centers to individual devices. Watch for future performance benchmarks comparing the version 2 model against the original 15-22 millisecond latency standard.