EmbeddingGemma 2 embeds text, images, video and audio under Apache 2.0

Google DeepMind's 740M EmbeddingGemma 2 maps text, code, images, video and audio into one 768-dimension space, runs on-device and moves from the Gemma license to Apache 2.0.

  • New release
  • 2 min read
  • Published

Google DeepMind released EmbeddingGemma 2 on 6 October 2026. It is a 740-million-parameter embedding model that places text, code, images, video and audio in one search space, and it is the first EmbeddingGemma under the Apache 2.0 license.

What is new

EmbeddingGemma 2 is built on the Gemma 4 architecture. It maps text (including source code), images, video and audio, alone or in combination, into a single 768-dimensional vector space, understands more than 100 languages and has an 8,192-token context window. Google says that is enough for about 5.5 minutes of audio, 29 images or 58 video frames. A voice memo can therefore find a video clip, and a sentence can find a photo.

The design is modular. A 270-million-parameter text and code model is the core, and you load a 170-million-parameter vision encoder and a 300-million-parameter audio encoder only if you need them, which gives 270M, 440M, 570M or 740M parameters. Matryoshka training lets you shorten the vectors from 768 to 512, 256 or 128 dimensions, which cuts storage by up to six times, and Google says quality holds up well down to 256 dimensions. Text inputs work best with short task prefixes, such as a search-query prefix.

Running it

The BF16 checkpoint is about 1.5 GB. Google says that, with quantization, the text-only model needs about 191 MB of RAM and the full multimodal model about 567 MB on a Pixel 11 Pro phone. Because it shares its tokenizer and audio encoder with Gemma 4, the two models can run together with a smaller combined memory footprint, which suits on-device use.

Google reports a score of 78.68 on the MTEB code benchmark, against 68.76 for EmbeddingGemma 1, and about the same multilingual text score as before, 61.36 against 61.15. It also says the model leads sub-1B multimodal embedders. Those are Google’s own results, so test it on your data.

License

The first EmbeddingGemma used the Gemma license. EmbeddingGemma 2 is Apache 2.0, with no revenue cap, hosting clause or territory limit, so it can go into a commercial product or a hosted service without a separate agreement.

Why it matters

Three open multimodal embedding families now sit side by side. Qwen3-VL-Embedding from Alibaba covers text, images, screenshots and video in 2B and 8B sizes, and Nemotron 3.5 Embed from NVIDIA covers text and document page images in 8B and 1B sizes, both with a context of 32K tokens. Qwen3-Embedding stays text-only. EmbeddingGemma 2 is the smallest of these and the only one that adds audio, but its 8,192-token window is shorter than the 32K of the others.

Vectors from different models are not interchangeable, so choosing one means re-indexing if you switch later. Qwen3-Embedding vs EmbeddingGemma 2 compares the two directly, and how to choose an embedding model walks through the decision.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
EmbeddingGemma 2 Google DeepMind 270M – 740M 8K Apache 2.0 (Permissive) 0.621 GB Laptop or phone
Qwen3-VL-Embedding Alibaba Qwen 2.13B – 8.14B 32K Apache 2.0 (Permissive) 4.9 GB Laptop or phone
Nemotron 3.5 Embed NVIDIA 1.55B – 8.38B 32K OpenMDW-1.1 (Permissive) 1.46 GB Laptop or phone
Qwen3-Embedding Alibaba Qwen 600M – 7.57B 32K Apache 2.0 (Permissive) 1.38 GB Laptop or phone
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone