Google releases multimodal EmbeddingGemma 2

Google DeepMind has released EmbeddingGemma 2. It follows last year's text-only EmbeddingGemma, which the company says has passed 20 million downloads. The new model is built on the Gemma 4 architecture, has 740 million parameters, and is released under the Apache 2.0 license.

The model is modular. Text-only workloads need as little as 270M parameters. Vision support adds a 170M encoder and audio support adds a 300M encoder. The context window is now 8K tokens, four times the original. Google says that is enough for up to 5.5 minutes of audio, 29 images or 58 video frames, or a mix of them. Matryoshka Representation Learning lets developers shorten the 768-dimension vectors to 512, 256 or 128 dimensions, which cuts storage by up to 6x.

On a Google Pixel 11 Pro with quantization, Google reports about 191MB of active RAM for text-only use and about 567MB for the full multimodal model.

Google says the model leads sub-1B multimodal embedders on MTEB Code and MAEB, the Massive Audio Embedding Benchmark. Its MTEB Code score rose from 68.76 to 78.68. Google says text performance stays in line with the first version. The announcement does not include full comparison tables.

EmbeddingGemma 2 shares a tokenizer and audio encoder with Gemma 4, so the two can run together in on-device RAG pipelines with a smaller combined memory footprint. Weights are on Hugging Face and Kaggle. Support is available at launch in transformers, sentence-transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LMStudio, Qdrant and Unsloth. Model Garden availability is listed as coming soon.