Google's EmbeddingGemma 2 brings multimodal search to phones, under Apache 2.0
Google DeepMind released EmbeddingGemma 2 on October 6, an open-weight embedding model that places text, code, images, video, and audio into one shared 768-dimensional vector space, letting a single on-device model power search and retrieval across media types that previously needed separate embedders. The model ships under the Apache 2.0 license, built on the Gemma 4 architecture, and follows an original EmbeddingGemma that Google says has passed 20 million downloads since its 2025 release.
One model, four modalities, selectively loaded
The release's core design choice is modularity. A 270-million-parameter text-and-code backbone handles the baseline workload; optional vision and audio encoders extend the model to its full 740 million parameters, and developers can load only the encoders a given application needs. A note-taking app might ship the text backbone alone, while a media-search tool could load the full multimodal stack. All configurations, regardless of which encoders are active, write into the same vector space, so a text query can retrieve a matching image or audio clip without a translation step between separate embedding systems. All modalities share an 8,192-token context window, four times the original EmbeddingGemma's.
Built for consumer hardware, not data-center retrieval
Google is positioning the model explicitly for on-device use. On a Pixel phone, the text-only configuration uses roughly 191 MB of active RAM; the full multimodal setup, quantized, runs in about 567 MB. The practical implication is that embeddings for personal photos, voice memos, and documents can be generated locally, making that content searchable without uploading it to a server. The model integrates with sentence-transformers, LiteRT and MediaPipe for on-device deployment, plus transformers, MLX, and Ollama, and supports fine-tuning for developers who want to adapt it to a narrower domain.
Where the benchmark gains concentrate
Google's own reported benchmarks show uneven gains across modalities. The largest jump is in code retrieval: the MTEB Code score rose 9.92 points, from 68.76 to 78.68. Multilingual text performance, by contrast, holds roughly level with the original EmbeddingGemma rather than improving, scoring 61.36 on MTEB multilingual v2. The newly added modalities post more moderate figures: 57.28 on MMEB v2 image retrieval, 67.84 on visual-document retrieval, 50.67 on video retrieval, 69.54 on MSEB sound retrieval, and 49.39 on MAEB audio. Google frames the model as the leading multimodal embedder under one billion parameters and claims it outperforms some specialist models more than twice its size, though neither technobezz.com's nor Unite.AI's coverage names specific competing models for that comparison, leaving the claim unverified against a named baseline.
Matryoshka representation learning does the compression work
The model uses Matryoshka Representation Learning, a training technique that front-loads the most important information in each vector so the vector can be truncated without a full re-encoding. Vectors can be cut from the native 768 dimensions down to 512, 256, or 128. Google reports that 256 dimensions retain about 95% of retrieval quality for image, video, and speech tasks, while 128 dimensions are sufficient for text-only workloads. The storage math is the more concrete payoff for developers: one million vectors at 768 dimensions cost roughly 1.5 GB, versus about 250 MB at 128 dimensions, a six-fold reduction that directly lowers both storage cost and lookup latency for any application built on top of the embeddings.
What's missing and what's next
Google's own documentation flags that the model is pre-trained only, with no safety tuning or post-training alignment applied, and that performance may vary across languages; it also recommends bfloat16 or float32 precision, since float16 can produce NaN values during inference. Weights are available now on Hugging Face and Kaggle, while listing in Google's Gemini Enterprise Agent Platform Model Garden is described as coming soon. For a site tracking which open releases get real adoption rather than announcement attention, EmbeddingGemma 2's test is whether the modular, selectively-loadable design and the Matryoshka compression actually translate into on-device deployments at the scale its predecessor reached, or whether multimodal complexity proves harder for third-party developers to integrate than a text-only embedder.