EmbeddingGemma 2: Benchmarks, Specs and How to Run It

§ 01What Google shipped#

An embedding model turns an input into a list of numbers so that similar things land close together. Search, retrieval-augmented generation (RAG), clustering and classification all run on that distance. What is new here is that one model handles five kinds of input in the same space. A single input can mix text with images, video and audio and still return one embedding, so a product described in text, two photos and a test video can be matched directly against a plain typed query. Google’s own examples are finding a video clip from a voice memo and searching hours of audio recordings from a text query.

The context window is 8K tokens, which Google calls “4x larger than EmbeddingGemma 1”. Per the announcement, that fits up to 5.5 minutes of audio, 29 images or 58 video frames in one input.

§ 02The benchmarks, from Google’s model card#

Google published scores for EmbeddingGemma 2 against its own predecessor only. The announcement claims “leading scores among sub-1B multimodal embedders” and says the model outperforms “some specialist models more than twice its size”, but neither the announcement nor the card names the models it was measured against. Every number below is Google’s, run on the full-precision checkpoint.

Code is the headline: a 9.92-point gain on MTEB Code, which Google ties to local codebase indexing, semantic code search and retrieval for coding agents. Multilingual text barely moved (61.36 against 61.15), so a text-only user who upgrades keeps the same retrieval quality and gains four times the context. The image, video and audio rows have nothing to compare against yet; they are first scores, and they will mean more once public leaderboards place named rivals beside them.

§ 03Smaller vectors: what each size costs#

EmbeddingGemma 2 is trained with Matryoshka Representation Learning (MRL): the leading numbers in each vector carry the most information, so a 768-number vector can be cut to 512, 256 or 128 numbers and still work. That is up to 6x less storage in a vector database. The cost shows up in the last step.

Google’s own reading: “Quality is close to lossless down to 256 dimensions.” Its developer guide says 256 keeps most of the full quality on text and code and “about 95% on image, video, and speech retrieval”. At 128 the multimodal score falls from 59.01 to 45.65, and the card says plainly that “128 dimensions degrades multimodal quality substantially”; it recommends 128 for text-only work. Queries and documents must use the same size: a 768-dimension query cannot be scored against a 128-dimension index.

§ 04Running it on a laptop or a phone#

The encoders are separate pieces, so a deployment loads only what it needs. All four configurations write into the same vector space, so a text-only index stays compatible if images are added later.

With quantization on a Google Pixel 11 Pro, Google measured about 191MB of active RAM for the text-only weights and about 567MB for the full multimodal model.

Every input draws on one 8,192-token budget, at fixed rates.

A smaller vision budget (Google allows 70 to 1,120 tokens per image) fits up to about 114 images or frames, at some cost in quality. Mixed inputs share the same budget, so adding one kind of input leaves less room for the others.

Weights are on Hugging Face and Kaggle; Google lists availability in the Gemini Enterprise Agent Platform Model Garden as coming soon. Day-one support covers sentence-transformers (version 6.1.0 or later), transformers, MLX, vLLM, llama.cpp, SGLang, Ollama, LM Studio, and transformers.js or WebGPU in the browser, plus Google’s own MediaPipe and LiteRT for apps and an Unsloth guide for fine-tuning. Because it shares Gemma 4’s text tokenizer and audio encoder, the two can run in one on-device pipeline with a smaller combined memory footprint: Gemma 4 reasons, EmbeddingGemma 2 retrieves.

Google’s developer guide shows the pattern for coding agents. It embedded the Hugging Face transformers codebase with the 270M text-only setup, then had an agent search it by similarity, using “Gemma 4 26B A4B with the Pi agent harness”.

§ 05What this means if you run agents#

Embeddings are the retrieval layer under most agent memory and RAG: an agent finds the right file, clip or past note by distance before it reasons about it. A model this small moves that index onto the laptop or phone where the data already sits, so private files can be searched without leaving the device. The code gain matters most to coding agents, which search a repository before they edit it. The open questions are the ones Google’s tables leave open: how it ranks against named rivals, and what the 256-dimension shortcut costs on your own data.

§ 06What we are watching#

• Named comparisons. Public MTEB, MMEB and MAEB leaderboard entries that put EmbeddingGemma 2 beside other sub-1B embedders; Google’s pages name none.
• Model Garden. Google lists Gemini Enterprise Agent Platform availability as coming soon.
• On-device defaults. Whether Google AI Edge apps (Gallery’s media search and video moment finder, and Foresight) become the reference way to ship it on Android.

§ 07The record#

As of October 6, 2026, 16:17 UTC: Google’s announcement on blog.google carries a publish time of 16:00 UTC. The Hugging Face repository google/embeddinggemma-2 was public before that: our news desk logged it at 15:42 UTC, and the repository shows a creation date of September 14 and a last change at 15:29 UTC on October 6. It lists the Apache 2.0 license and showed 364 downloads at our read. All benchmark figures are Google’s, from the model card.

§ 08Sources#

• Google, EmbeddingGemma 2: an open, lightweight multimodal embedding model, Sahil Dua and Henrique Schechter Vera, Google DeepMind, October 6, 2026; read 16:16 UTC.
• Google DeepMind, EmbeddingGemma 2 model card on Hugging Face, read October 6, 2026; also on Google AI for Developers.
• Google Developers Blog, EmbeddingGemma 2: The Developer Guide, Maarten Grootendorst and Ian Ballantyne, October 6, 2026.
• Kaggle, google/embeddinggemma-2, read October 6, 2026.
• Google, Introducing EmbeddingGemma, the first version.