Google's EmbeddingGemma 2 Embeds Text, Images, Audio and Video in 740M Parameters, but English Retrieval Slipped

Google DeepMind released EmbeddingGemma 2 on October 6, a 740M-parameter model that maps text, code, images, video and audio into one 768-dimension vector space. The weights are on Hugging Face and Kaggle, and Google's launch post gives the licence as Apache 2.0.

The post is written by Sahil Dua and Henrique Schechter Vera, both Research Engineers at Google DeepMind. It says the first EmbeddingGemma passed 20 million downloads.

What EmbeddingGemma 2 is made of

The 740M total splits into a 270M text model, a 170M vision encoder and a 300M audio encoder. The model card says developers can load only what they need: 270M for text, 440M with vision, 570M with audio.

The context window is 8,192 tokens shared across modalities, up from 2,048 in the first model. The card prices an image at 280 tokens, a video frame at 140 and a second of audio at 25. That works out to about 29 images, 58 frames or 327 seconds of audio per input.

The card describes the model as pre-trained only. It has no post-training alignment and no output moderation, and the card tells developers to build their own safeguards.

The scores Google reports

The headline gain is on code. MTEB Code NDCG@10 rose from 68.76 to 78.68, according to the card. On MTEB multilingual v2 the gain is 61.15 to 61.36.

On MTEB English v2 the card shows 68.46, and the review site eesel.ai put the first model's score at 69.67, a slip of about one point. Google's post does not mention that figure.

Google calls the model best in class among sub-1B multimodal embedders on MTEB Code and MAEB, an audio benchmark. The eesel.ai review found no head-to-head table on the model card behind that claim. All figures are vendor-supplied.

Truncating the vectors costs more in vision than in text

The model supports Matryoshka truncation, which cuts the 768 dimensions to 512, 256 or 128. Google's launch post says that saves up to 6x in vector storage, while its AI Edge post says up to 8x. The two Google pages disagree and neither explains the gap.

Dimensions	MTEB multilingual	MTEB Code	MMEB v2 overall
768	61.36	78.68	59.01
512	61.17	77.24	58.38
256	60.41	76.18	56.24
128	57.89	71.41	45.65

The card says quality is close to lossless down to 256 dimensions and that 128 suits text-only work. The 128-dimension row shows why: the multimodal benchmark MMEB v2 falls by 13.4 points, against 3.5 for multilingual text.

The card also warns against float16, which it says can produce NaN values or degraded embeddings without an error. Use bfloat16 or float32, and renormalise after truncating.

On-device numbers and what is missing

Google's AI Edge team measured about 191 MB of active RAM for the text-only weights and about 567 MB for the full multimodal model on a Pixel 11 Pro. The post does not state the quantisation for those two figures.

Vision embedding took 37.3 ms, or 26.9 images a second, on a MacBook M5 Pro GPU at a budget of 70 vision tokens per image. The per-device CPU, GPU and NPU table is an image, and the post gives no text-embedding latency.

The model runs in sentence-transformers, Transformers, vLLM, SGLang, MLX, llama.cpp, Ollama and LM Studio. Google lists Android support through ML Kit and availability in its Model Garden as not yet shipped.

Independent testing is thin

The eesel.ai review found no formal independent benchmarks. It cited informal Hacker News reports, including one in which Google's own zero-shot example scored a refund request 0.22 for being financial.

The model sits in a week of small open releases. Cloudflare's Clef targets a different task with the same Apache 2.0 licence, and Alibaba's Qwen3.8-27B is the larger text model a retrieval pipeline might pair with it.

The card's licence link points to a document titled gemma_4_license, while the page tag and Google's post both say Apache 2.0. The text of that document was not read for this report.

Google has given no date for the ML Kit release beyond the phrase "in the coming weeks".