EmbeddingGemma 2 Maps Text, Video and Audio Into One 768-Dimension Space

Google DeepMind released EmbeddingGemma 2 on Oct. 6, an open model of about 740 million parameters that maps text, code, images, video and audio into a single 768-dimension vector space. The weights are listed on Hugging Face under the Apache 2.0 licence.

The model card, published under the google organisation on Hugging Face, shows about 2,000 likes and 124,930 downloads over the last month. The first EmbeddingGemma handled text only.

What the 740 million parameters are made of

The card breaks the total into four parts: a 130 million-parameter backbone, a 140 million-parameter embedder, a 170 million-parameter vision encoder and a 300 million-parameter audio encoder. The text path is the first two, 270 million parameters, and LetsDataScience reports that developers can load only the encoders a task needs.

The model has 24 layers, a vocabulary of 262,144 tokens and a sliding attention window of 1,024 tokens. Its context length is 8,192 tokens, shared across all modalities. Native output is 768 dimensions, and the card lists truncation to 512, 256 and 128 dimensions.

What the scores show, and what they cannot

Google's card compares EmbeddingGemma 2 with its predecessor on only two benchmarks. Text retrieval across languages barely moved. Code retrieval did.

Benchmark (768 dimensions)	EmbeddingGemma 2	EmbeddingGemma 1
MTEB multilingual v2	61.36	61.15
MTEB code v1	78.68	68.76

1. EmbeddingGemma 168.76 points
2. EmbeddingGemma 278.68 points

Source: EmbeddingGemma 2 model card, Hugging Face, accessed 2026-10-13

The code gain is 9.92 points, which matches the card's roughly 14 percent claim. Every figure here is Google's own. The multimodal scores, 57.28 on MMEB v2 Image, 67.84 on MMEB v2 VisDoc and 50.67 on MMEB v2 Video, come with no earlier EmbeddingGemma baseline because the first model could not take those inputs. The card's own verdict, "among the strongest multimodal embedding models under 1B parameters," is a vendor claim with no named rival and no independent run behind it.

What the card tells developers to do

Several instructions on the card will matter to anyone who deploys it. It says to run the model in bfloat16 or float32, because float16 can produce NaN or degraded embeddings. Text inputs benefit from task prefixes such as task: search result | query:, while images, video and audio take none. Vision budgets range from 70 to 1,120 soft tokens per image.

The card says truncation to 256 dimensions is close to lossless and that 128 dimensions suits text-only work. LetsDataScience advises developers to test ranking quality on their own data before trusting any truncation.

What is missing

The card describes a pre-trained model with no post-training alignment, safety tuning or output moderation, and leaves safeguards such as retrieval filtering to the developer. The fetched card gives no release date, and it shows no gating notice. The card as fetched lists no memory or latency figure, so claims of what the model needs on a phone rest on the reader running it.

The Terminal has covered what memory capacity with no bandwidth figure leaves unanswered for inference hardware, and the same gap applies to on-device embedding. Local models of this size also feature in Anthropic's red-team warning about small models on cheap devices. Independent MTEB and MMEB runs on the Hugging Face weights are the next numbers to look for.