News · 2026-10-06
Google ships EmbeddingGemma 2: 740M-parameter open multimodal embeddings for on-device RAG
Google DeepMind released EmbeddingGemma 2 on 2026-10-06: a 740M-parameter embedding model on the Gemma 4 architecture under Apache 2.0 that maps text, code, images, video, and audio into one shared 768-dimensional space. Text-only mode runs 270M parameters, and quantized RAM is about 191MB text-only or 567MB full on a Pixel 11 Pro. The 8K context is four times the first model's, MTEB Code rises 9.92 points from 68.76 to 78.68, and Matryoshka truncation cuts vectors to as few as 128 dimensions for up to 6x storage savings. Weights are on Hugging Face and Kaggle with day-one support in Ollama, llama.cpp, LM Studio, and vLLM.
What it means for buyers
Local RAG stacks get a drop-in multimodal retriever that pulls and runs in Ollama or LM Studio today, so on-device search over docs, screenshots, and recordings no longer needs a hosted embedding API. The code-score jump is pitched at coding-agent retrieval, and per-vendor vectors with Gemma 4 keep a local pipeline inside one memory budget. Benchmark numbers are Google's own against its predecessor only.
Where to check it
Sources
- Google DeepMind: EmbeddingGemma 2, an open lightweight multimodal embedding model — vendor primary: parameters, license, benchmarks, RAM figures, checked 2026-10-07
- Hugging Face: google/embeddinggemma-2 weights — weights availability, checked 2026-10-07