EmbeddingGemma 2

Released October 6, 2026, EmbeddingGemma 2 is Google DeepMind's open multimodal embedding model under Apache 2.0, folding five embedding jobs — text, code, images, audio, and video — into one 740M-parameter model that maps everything into a single 768-dimensional vector space. It is modular: text-only workloads need just 270M parameters, with optional 170M vision and 300M audio encoders loaded only when needed, so a phone app can run retrieval on ~191MB of RAM (567MB for the full stack, quantized). Matryoshka Representation Learning allows truncating vectors to 512, 256, or 128 dimensions, cutting vector-database storage up to 6x, and the 8K-token context handles about 5.5 minutes of audio or 29 images per pass. Google reports a 9.92-point MTEB Code jump over its predecessor (68.76 to 78.68).

Limitations: Brand new (October 2026) with benchmarks so far company-reported. The unified-space convenience means per-modality quality may trail dedicated specialist embedders at larger sizes. On-device deployment still requires the MediaPipe/LiteRT/transformers.js toolchain, and re-embedding existing indexes to adopt it is a real migration cost for production RAG systems.

Google-reported: MTEB Code 78.68 (+9.92 vs v1); leads sub-1B multimodal embedders