Updated

EmbeddingGemma 2

Google launched EmbeddingGemma 2, an open multimodal embedding model that maps text, code, images, video, and audio into a shared embedding space. Its 740M-parameter weights are available under Apache 2.0 for on-device search and retrieval.

By No Code MBA · Facts checked

Provider
Google
Announcement date
Status at last check
Generally available
Model family
EmbeddingGemma
Preceding model
EmbeddingGemma
Capabilities
Text, Image, Audio, Video, Coding, Embeddings

What changed

EmbeddingGemma 2 extends the text-only EmbeddingGemma to text, code, images, audio and video in one embedding space, with 740M parameters built on the Gemma 4 architecture. Text-only use needs as little as 270M parameters, with optional 170M vision and 300M audio encoders. Context grows to 8K tokens, 4x the first version, and output vectors can be truncated from 768 to 512, 256 or 128 dimensions. Google reports its MTEB Code score rose from 68.76 to 78.68.

Access and availability

Weights are on Hugging Face (google/embeddinggemma-2) and Kaggle under Apache 2.0, with on-device builds in the LiteRT Community on Hugging Face. It works with sentence-transformers, transformers.js, Ollama, llama.cpp, vLLM and other runtimes, and Google says Model Garden support is coming soon.

Official access documentation

What you can build

Compare local retrieval with your current embedding model using the same text, code, image, audio, and video queries; review relevance, latency, and device memory before adopting it.

Limitations

Google's model card notes that 128-dimension vectors degrade multimodal quality substantially, float16 can produce NaN or degraded embeddings, and skipping the recommended task prefix reduces quality. The model has no safety tuning and performance may vary across languages.

Sources and release history