EmbeddingGemma 2 Packs Text, Code, Images, Video and Audio Into One 740M-Parameter Embedder

Google DeepMind released EmbeddingGemma 2 on October 6, 2026. The open embedding model maps text, code, images, video, and audio, alone or mixed together, into a single 768-dimensional vector space. It is built on the Gemma 4 architecture, has 740 million parameters in its full form, and is released under the Apache 2.0 license. The weights are on Hugging Face and Kaggle, and Google says availability in its Model Garden is coming soon. According to Google, the first EmbeddingGemma, a text-only model released last year, has passed 20 million downloads. Developers have used it for on-device search and privacy-focused retrieval-augmented generation (RAG). The new version keeps that size class and extends it to media.

Four encoder setups, one vector space

The model is modular. Per the model card, the text path takes 270M parameters, a 130M transformer backbone plus a 140M embedder, and handles text and code. A 170M vision encoder covers images, visual documents such as PDFs and slides, and video frames. A 300M audio encoder takes raw audio directly. Developers load only what they need: 270M for text only, 440M with vision, 570M with audio, or the full 740M.

All four setups load from the same checkpoint and project into the same space. Google’s developer guide says a query embedded with the text-only setup can be matched directly against documents embedded with the full model. A team that indexes text first and adds images or audio later does not have to recompute its existing vectors.

The context window is 8,192 tokens, four times that of the first EmbeddingGemma, and every modality draws from that one budget. An image costs 280 tokens by default, a video frame 140, and a second of audio 25. That works out to about 29 images, 58 video frames, or roughly five and a half minutes of audio in a single input. Video is sampled at one frame per second by default, and audio should be 16 kHz mono. Inputs can also be interleaved: placeholder tokens in the text mark where each image, clip, or recording goes, and the call returns one embedding for the whole bundle, such as a product listing with a photo and a demo video.

Code search gets the biggest lift

The headline improvement is code. On MTEB Code, the score rises from 68.76 to 78.68, a gain of 9.92 points, while multilingual MTEB v2 text holds roughly steady at 61.36 against 61.15. Google pitches the code result at local codebase indexing and at retrieval for coding agents. Its developer guide shows a demo that indexes the Hugging Face transformers codebase with the 270M setup and searches it through an agent running Gemma 4 26B A4B on the Pi harness.

The model card adds first scores for the new modalities: 64.64 on MIEB lite and 57.28 on MMEB v2 image tasks, 67.84 on visual documents, 50.67 on video, 69.54 on MSEB retrieval, and 49.39 on MAEB for audio. Google calls these leading results among sub-1B multimodal embedders. These are Google’s own figures, measured on the full-precision checkpoint.

Storage, memory and a float16 warning

EmbeddingGemma 2 is trained with Matryoshka Representation Learning, so vectors can be cut from 768 dimensions to 512, 256, or 128. The developer guide gives the storage math: a million 768-dimensional vectors in bfloat16 take about 1.5 GB, while 128 dimensions need about 250 MB. The cost shows up first in media. At 256 dimensions, the model keeps most of its full quality on text and code and about 95% on image, video, and speech retrieval. At 128 dimensions, text and code keep around 90%, but multimodal retrieval falls to around 75%. Google recommends 128d mainly for large text-only indexes. Truncated vectors must be re-normalized, and queries and documents must use the same dimension.

On the phone, Google says the quantized model needs about 191 MB of active RAM for the text-only weights and about 567 MB for the full multimodal model on a Pixel 11 Pro. The model card has one warning in particular: do not run the model in float16. Its activations exceed float16’s range, and the model then returns NaN or silently degraded embeddings instead of an error. The card recommends bfloat16, or float32 on most CPUs.

Tooling on day one

Support covers sentence-transformers (v6.1.0 or later for multimodal input), transformers, transformers.js, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, and LiteRT, with Qdrant named for vector storage and Unsloth for fine-tuning. Text inputs use short task prefixes for search, question answering, fact checking, code retrieval, classification, clustering, and similarity. The card lists a training-data cutoff of January 2025. Because it is a pre-trained embedder with no output-level moderation, Google says retrieval filtering and fairness testing are the deployer’s responsibility.