EmbeddingGemma 2

EmbeddingGemma 2 is an open, multimodal embedding model built by Google DeepMind. It maps text (including code), images, video and audio — as well as combinations of them — into a single, unified 768-dimensional vector space. The model has 740M total parameters: a 270M text backbone (130M transformer + 140M embedder) complemented by selectively loadable vision (170M) and audio (300M) encoders.

It is designed to run on consumer hardware such as mobile devices and laptops, targeting low-latency on-device use cases: semantic search, retrieval-augmented generation (RAG), classification and clustering. The context window is 8,192 tokens, enough to process minutes of audio or video.

Key features

• Native multimodality: four modalities (text, images, video, audio) in a single 768-dimensional space.
• Multilinguality and code: understands 100+ languages and achieves about a 14% improvement on code tasks relative to its predecessor.
• Matryoshka Representation Learning (MRL): native truncation of embeddings to 128, 256 and 512 dimensions, for up to roughly 6x reduction in vector storage cost with minimal quality impact.
• Flexible footprint: selectively loadable encoders (text-only 270M, text+image 440M, text+audio 570M, full multimodal 740M).

EmbeddingGemma 2 builds on the architecture and improvements of the Gemma 4 family and is released under the open Apache 2.0 license. Recommended inference precision is bfloat16 or float32 — using float16 yields NaN or silently degraded results.