TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Google announced EmbeddingGemma 2 on Oct. 6, 2026, an open-weight model that maps text, code, images, audio and video into a shared embedding space. Google says it has 740 million parameters, supports local inference and is available under the Apache 2.0 license; its benchmark and device-performance claims are from the company.
Google announced EmbeddingGemma 2 on Oct. 6, introducing a 740-million-parameter embedding model that can represent text, code, images, audio and video in a shared space for search and retrieval. Google says the model is designed for on-device use and is released under the commercially permissive Apache 2.0 license, giving developers an option for cross-modal search that can run locally.
Google says EmbeddingGemma 2 is built on the Gemma 4 architecture and uses technology from its Gemini Embedding models. Rather than generating answers, an embedding model converts information into numerical vectors that can be compared to retrieve related items. In the new model, a developer could use a text query to find relevant audio or video content, according to Google. The company says the model can process combinations of modalities in one system.
The model is modular: Google lists a 270-million-parameter text-only configuration, with optional vision and audio encoders of 170 million and 300 million parameters, respectively. It supports an 8,000-token context window, which Google says can cover up to 5.5 minutes of audio, 29 images or 58 video frames, or combinations of those inputs. The limits depend on the input and its representation; the announcement does not describe every implementation detail.
Google reports that EmbeddingGemma 2 improves its predecessor’s score on the MTEB Code benchmark from 68.76 to 78.68, a 9.92-point gain, while maintaining strong multilingual text performance. It also says the model scores well on multimodal benchmarks, including MTEB Code and the Massive Audio Embedding Benchmark. These are company-reported evaluations; the announcement points readers to the model card for full metrics.
Local Search Across Media Types
EmbeddingGemma 2 targets developers who want to search across different kinds of files without first sending their contents to a cloud service. If the model and search index run locally, an app could retrieve relevant material from a device’s recordings, images or documents while working offline. Google says local embedding can reduce latency and support privacy-focused retrieval-augmented generation, in which a system retrieves relevant sources for a generative model.
The release also matters for developers working under memory and storage constraints. Google says the model can reduce vector size from 768 dimensions to 512, 256 or 128 using Matryoshka Representation Learning, with up to sixfold lower storage use. It reports active RAM needs on a Pixel 11 Pro of about 191 megabytes for quantized text-only weights and 567 megabytes for the full multimodal model. Those figures are Google’s measurements for the stated device and setup, not a guarantee of performance across hardware.
The Apache 2.0 license allows commercial use under its terms, which may make the model useful to teams that need to inspect, adapt or deploy model weights in their own products. The license does not by itself establish how well the model will perform in a particular app, or resolve developers’ responsibilities for the data they index.
Products Worth Considering
Apple Watch Series 12 GPS 46mm
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Text Embeddings to Multimodal
Google introduced the original EmbeddingGemma last year as a lightweight model for text embeddings. In its announcement, the company said that model had received more than 20 million downloads and was used for on-device search and retrieval-augmented generation. The new release extends the product line from text to code, images, audio and video.
Google says EmbeddingGemma 2 shares a text tokenizer and audio encoder with Gemma 4, which may allow developers to combine the models in one local pipeline with a lower combined memory footprint. It has not provided a single memory figure for every paired configuration. The company lists several ways to use the model, including Google AI Edge MediaPipe and LiteRT for on-device deployment, as well as Hugging Face and Kaggle for model downloads. Availability in Gemini Enterprise Agent Platform’s Model Garden is described as coming soon.
““EmbeddingGemma 2 is the most capable model for on-device multimodal embeddings.””
— Sahil Dua and Henrique Schechter Vera, Google DeepMind research engineers, in Google’s announcement
Independent Tests and Device Results
The announcement does not provide independent verification of Google’s benchmark rankings or comparisons with larger and specialist models. Full evaluation details are linked in the model card, but results can depend on test sets, scoring methods and implementation choices. The report also does not give a broad device-by-device performance assessment, battery impact or latency figures.
Google’s RAM estimates refer to a quantized model on a Pixel 11 Pro; the announcement does not specify enough detail to assume those figures apply to other devices or configurations. It also remains unclear how well the model handles particular languages, media formats and noisy recordings in real-world applications. The stated Apache 2.0 license covers the model release, but the announcement does not settle questions about rights or consent for user-provided content indexed by an app.
Model Access and Developer Testing
Developers can download EmbeddingGemma 2 weights from Hugging Face and Kaggle, according to Google. The company says optimized versions are available through the LiteRT Community on Hugging Face and points to documentation for inference, deployment and fine-tuning. Its announcement lists MediaPipe, LiteRT, transformers.js and WebGPU among deployment routes, alongside tools such as Transformers, sentence-transformers, MLX, vLLM and llama.cpp.
Gemini Enterprise Agent Platform Model Garden availability is expected later, but Google did not give a date. Independent evaluations and hands-on tests across different hardware will help establish whether the reported quality and memory figures hold for developers’ intended workloads.
Key Questions
What is EmbeddingGemma 2?
It is Google’s 740-million-parameter embedding model for representing text, code, images, audio and video in a shared vector space, so related items can be retrieved through search.
Can it run without an internet connection?
Google designed it for on-device inference and says it can support offline search and retrieval. Whether an app works fully offline depends on its deployment and other services it uses.
What license does Google provide?
Google released the model under the Apache 2.0 license, which permits commercial use subject to the license’s terms.
How much memory does it require?
Google reports about 191 MB of active RAM for quantized text-only weights and about 567 MB for the full multimodal model on a Pixel 11 Pro. Requirements on other devices may differ.
Where can developers get the model?
Google says weights are available on Hugging Face and Kaggle. Availability in Gemini Enterprise Agent Platform Model Garden is listed as coming soon, without a specific date.
Source: hn
