What happened
Google DeepMind released EmbeddingGemma 2 on October 6, 2026. It is an open embedding model that maps text, code, images, video and audio into a single shared vector space, designed to run on phones, laptops and other local hardware. The original EmbeddingGemma was text-only and, according to Google, passed 20 million downloads.
What’s new
- Size: 740M parameters in total. Text-only use needs about 270M; the vision (170M) and audio (300M) encoders are optional.
- Base: built on Gemma 4, sharing its text tokenizer and audio encoder.
- Context: 8K tokens, four times the original. Google says that covers roughly 5.5 minutes of audio, 29 images or 58 video frames, or mixed combinations.
- Flexible dimensions: Matryoshka Representation Learning lets you truncate 768-dimension vectors to 512, 256 or 128 for up to 6x storage savings.
- Memory on device (quantized, Pixel 11 Pro): about 191MB active RAM for text-only and about 567MB for full multimodal use.
- Results: MTEB Code improves from 68.76 to 78.68 versus the original model; text performance matches the original’s multilingual results. Audio, image and video scores are in the model card.
- License: Apache 2.0, which allows commercial use.
Why it matters
Retrieval-augmented generation and semantic search usually need separate embedding models for text, images and audio, plus a server to run them. A single small model with a shared space lets an app search photos with a text query, find a moment in a voice recording, or index mixed documents locally, without sending data to a cloud API. That is useful for privacy-sensitive apps, offline use and cutting embedding costs.
The Matryoshka option is practical for large indexes: storing 128-dimension vectors instead of 768 reduces vector database size substantially, at some cost in accuracy that teams should measure on their own data. Because the model is small and permissively licensed, it can also be fine-tuned on domain data, such as product catalogs or internal documents, without licensing restrictions on commercial deployment.
How to try it
- Download the weights from Hugging Face (google/embeddinggemma-2) or Kaggle. Availability in Google’s Model Garden is listed as coming soon.
- For servers, load it with sentence-transformers, transformers, vLLM, SGLang, llama.cpp (GGUF), Ollama, LM Studio or MLX.
- For on-device apps, use the LiteRT-optimized versions with MediaPipe or LiteRT, or transformers.js and WebGPU in the browser.
- Google points to Qdrant for vector storage and Unsloth for fine-tuning.