Google ships EmbeddingGemma 2, 744M unified multimodal embedder

Google's EmbeddingGemma team released a 744M-parameter model that encodes text, images, audio, video and combinations of them into a single 768-dimension embedding space. The paper on Hugging Face, posted on October 28, says the model "establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger" across comprehensive multimodal benchmarks.

The scores it leans on: 68.46 on MTEB(English, v2), 78.68 on MTEB(Code), and 69.54 MRR@10 on native-audio retrieval. The budget splits across a 271M text backbone, a 167M vision encoder and a 305M audio encoder, feeding one shared 768-dimension output that Matryoshka Representation Learning lets callers truncate down to 128.

On-device numbers come from a Google Pixel 11 Pro: 191 MB peak memory for the text-only Q4_C quantized variant, 447 MB with vision, 567 MB for the full multimodal stack. Model quality survives the squeeze. MTEB(English) actually ticks up a hair to 68.49 at int4 Q4_C.

One ablation stands out. Native audio understanding scores 69.54 overall against 50.95 for Google's own prior EG1 ASR-plus-text pipeline, and the gap widens cross-language (67.68 versus 41.87). The authors describe the design as supporting "selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights," which is why text-only shops can skip the heavy towers entirely at load time.

The drop lands amid a run of open-weights releases we have been tracking all week on the open-source feed.