EmbeddingGemma 2 brings text, image, audio, and video into one small embedding model

Google DeepMind has released EmbeddingGemma 2, the successor to last year's text-only EmbeddingGemma. The new model maps text, code, images, video, and audio into a single shared embedding space. It has 740M parameters, uses the Gemma 4 architecture, and ships under Apache 2.0. The headline isn't a new frontier result. It's that one small model now covers retrieval across every modality, and it fits on a phone.

What shipped

The announcement describes a model built to fit a device rather than to top a leaderboard:

• Modular weights. Text-only workloads need as little as 270M parameters. Add a vision encoder (170M) and an audio encoder (300M) only if you need them.
• Matryoshka outputs. 768-dimensional vectors can be truncated to 512, 256, or 128 dimensions, which DeepMind says cuts storage by up to 6x.
• Small memory footprint. Quantized on a Pixel 11 Pro, DeepMind reports roughly 191MB of active RAM for the text-only weights and about 567MB for the full multimodal model.
• 8K context. Four times the first model's window, enough for about 5.5 minutes of audio, 29 images, or 58 video frames.
• Code quality. MTEB Code rises from 68.76 to 78.68, with multilingual text performance held at the first model's level.

You pay only for the modalities you load, and you choose the vector size at index time.

Why it matters

In production, my objection to multimodal retrieval has rarely been quality. It has been the cost of running several systems: a text embedder, a CLIP-style model for images, and a speech-to-text step so audio can be searched as text. Each has its own vector space and index, and joining results across them is a ranking problem you solve yourself. A single shared space removes that layer. A text query can retrieve a video clip without going through a caption first.

The modular design matters as much. Most deployments I've worked on are mostly text, with a few images or recordings on the side. Loading the 270M core by itself means the multimodal option doesn't make the common case more expensive. Matryoshka truncation does similar work on a device: 128 dimensions instead of 768 can decide whether an index fits in a phone's spare memory, and you pick the size at index time after testing recall on your own data.

The pairing with Gemma 4 is also practical. DeepMind says the two models share a text tokenizer and audio encoder, so running both together uses less memory than running them separately. In on-device RAG, where the embedder and the generator compete for the same RAM, that saving is real.

My read

This is an incremental release, and that's fine. DeepMind frames its claims as quality per parameter: it says the model leads the sub-1B class on MTEB Code and MAEB, and matches or beats "many larger models" on text, vision, and audio. I'd treat that as a statement about size class rather than absolute rank, and compare the model card's tables with results on my own retrieval set before replacing anything. The first things I'd test are recall at 128 and 256 dimensions on domain data, and whether text-to-text scores systematically outrank text-to-image scores in a mixed-media index.

Ecosystem support is broad from day one: weights on Hugging Face and Kaggle; serving through transformers, sentence-transformers, vLLM, llama.cpp, Ollama, and MLX; on-device paths through MediaPipe and LiteRT. For privacy-sensitive or offline retrieval, EmbeddingGemma 2 is now the baseline to beat. The full details are in DeepMind's announcement.