One compact model embeds text, images, audio and video together

The authors report that a 744-million-parameter model often matches or exceeds larger and specialist embedding models across seven benchmark suites.

Big Tech

Henrique Schechter Vera · Sahil Dua · Jovan Andonov · Sara Smoot · Sebastian Russo · Ivan Llanos · +67 more

Google

Research Digest·28 October 2026·2 min read

Illustration: One compact model embeds text, images, audio and video together

In one line

EmbeddingGemma 2 unifies text, images, audio, and video into a single embedding space, outperforming models two to twelve times larger.

Before you read this

Two ideas this paper builds on

contrastive alignmentA method that learns embeddings by minimizing distance between positive pairs and maximizing distance between negative pairs.36 papers →quantization-aware trainingTraining a neural network with simulated quantization in the forward pass so its weights and activations adapt to low-precision formats, improving deployment accuracy.15 papers →

Read it as

Many AI systems need to understand different types of data like text, pictures, sound, and video together. Existing methods either use separate tools for each type, which adds cost and misses fine details, or rely on enormous models that are too expensive for many real-world uses.

EmbeddingGemma 2 is a single model that represents any combination of these data types as a single vector, a fixed-length list of numbers that captures meaning. It uses separate encoders for each modality that feed into a shared transformer, so it can run as a text-only or multimodal model. The training used a multi-stage curriculum and simplifications that improved efficiency.

Despite its modest size, EmbeddingGemma 2 matches or exceeds the quality of specialized models and models up to twelve times larger. This makes high-quality multimodal understanding practical for on-device and latency-sensitive applications.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

multimodal embedding curriculum learning contrastive alignment knowledge distillation embedding truncation quantization-aware training