Google DeepMind released EmbeddingGemma 2, a 740-million-parameter open embedding model that maps text, code, images, video, and audio into a single 768-dimensional vector space, under an Apache 2.0 license1,7. The model is small enough to run on a smartphone2, built on the Gemma 4 architecture3.
The release extends the EmbeddingGemma line beyond text, which was all the first version handled when Google introduced it in September 2025. Images, audio, and video now share one embedding space with text.
Architecture and footprint
EmbeddingGemma 2 combines a 270-million-parameter text backbone (130M transformer plus 140M embedder) with modular vision (170M) and audio (300M) encoders. Developers can load only the modalities required for a given use case, keeping the active parameter count below the 740M ceiling6. The model carries an 8K token context window and can process minutes of audio or video.
Native Matryoshka Representation Learning allows truncated embeddings at 128, 256, 512, and 768 dimensions, enabling up to a 6x reduction in vector storage costs with what Google described as minimal impact on quality. Lightweight text instruction prefixes optimize embeddings for tasks such as search, classification, clustering, and semantic similarity.
EmbeddingGemma 2 supports 100-plus languages and achieves a roughly 14% improvement on code tasks relative to its predecessor.
On-device and in-browser deployment
Designed for consumer hardware including mobile devices and laptops, EmbeddingGemma 2 delivers low-latency semantic representations for on-device search, retrieval-augmented generation, classification, and clustering. A community demo already runs the model locally in-browser on WebGPU4. The model also has llama.cpp support.
ANALYSIS A shared 768-dimensional embedding space across four modalities enables cross-modal retrieval: a text query can surface a relevant video clip, audio segment, or code snippet without modality-specific pipelines.
At 740 million parameters under a permissive open-source license, EmbeddingGemma 2 targets the retrieval and indexing layer that feeds downstream generation rather than general reasoning or agentic workloads.
The modular encoder design, combined with Matryoshka truncation, gives developers two independent levers to reduce serving cost: drop unused modalities and shorten vector dimensions.
The Hacker News post for the Google blog announcement drew 222 points and 29 comments as of October 6.