Google released EmbeddingGemma 2 on Tuesday, October 6, with open weights for a model designed to search across text, code, images, video and audio. That broader reach could give developers a useful new option, but it is not evidence that the model retrieves relevant results reliably across every format. We believe open models deserve trust only when their performance can be checked.
The weights are available on Hugging Face and Kaggle under the Apache 2.0 license, Google said. Availability through Gemini Enterprise Agent Platform Model Garden was still described as coming soon. Google’s launch announcement makes a real distinction between what developers can access now and what has yet to arrive.
Open weights widen the choice
An embedding model turns information into numerical vectors that let software find items with similar meanings. Google says EmbeddingGemma 2 maps text, code, images, video and audio into one shared space, extending the original EmbeddingGemma, which Google DeepMind introduced as a text-embedding model in September 2025.
That design could make it easier to search a collection that mixes documents, pictures and recordings without using a separate embedding model for each format. But the convenience is a design goal, not proof that relevant results will surface consistently in a working product.
Open weights add another kind of practical value: developers can choose how to deploy a model instead of depending only on a hosted service. Commentator Simon Willison has argued that this matters for embedding systems because switching providers can require recalculating stored vectors. Having weights available could offer a fallback if a hosted provider stops offering a model, he wrote. The original EmbeddingGemma publication and Willison’s discussion help show why openness matters beyond the license label.
That flexibility is worth welcoming. It does not settle whether the model is good enough for a particular search job, and the license alone tells users nothing about the quality of its results.
A longer list of inputs is not a quality score
Google’s model card describes EmbeddingGemma 2 as a 740-million-parameter model and reports scores across text, code, image, visual-document, video and audio benchmarks. Those figures are Google’s results; no broad independent evaluation has confirmed the model’s cross-modal performance in the sources reviewed.
The company reports a score of 78.68 on MTEB Code, up from 68.76 for EmbeddingGemma 1. On the multilingual MTEB benchmark, its reported score moves from 61.15 to 61.36. These task-specific measurements give developers something to investigate, but they are not a guarantee that search will work well on a company’s own files, languages or recordings. The model card contains the reported scores and setup.
Google DeepMind research engineers Sahil Dua and Henrique Schechter Vera describe it as “the most capable model for on-device multimodal embeddings.” That is the developer’s assessment, not an independently established ranking. The Frontier’s analysis of Google’s published chart also noted that a sub-1-billion-parameter competitor was plotted above EmbeddingGemma 2 on an audio score. The Frontier did not independently rerun that benchmark, so its observation is a useful caution about the comparison, not a substitute evaluation.
Google also reports that a quantized version used about 567 MB of active RAM for the full multimodal model on a Pixel 11 Pro, compared with about 191 MB for text-only weights. Those are Google’s measurements on that device, not a promise about memory use on other phones or in a complete application. On-device use is a meaningful aim; its costs and performance still need testing in context.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
Make the evidence reproducible
There is some outside testing, but it answers a narrower question. Vector-search company Qdrant tested compression with EmbeddingGemma 2 across five text-retrieval datasets. It reported that quantized full-size vectors retained 99% of baseline retrieval quality while using 30 times less vector RAM; 256-dimensional vectors with rescoring retained 94.5% while using 77 times less vector RAM.
Those results compare against exact search with full-size float32 vectors, and Qdrant says its memory figures exclude graph indexes, payloads and original vectors kept for rescoring. More importantly, the experiment concerns text retrieval and storage trade-offs. It does not verify how accurately EmbeddingGemma 2 retrieves images, video or audio. Qdrant itself advises developers to test relevance and latency on their own data in its published results.
The strongest case for a broad, open model is not that one system must beat every specialist. It is that developers can try a flexible option, inspect its behavior and decide whether it fits their needs. But that case works only if they can distinguish a company’s benchmark claim from a result others can reproduce.
We recommend that Google and other model publishers support independent, modality-specific evaluations with clear methods and reproducible settings. Developers adopting the model should test retrieval accuracy, latency and memory use against their own collections before building around it. Broader access is progress; reliable evidence is what lets users decide whether that progress is useful.