EmbeddingGemma 2

After quantization, EmbeddingGemma 2 requires only approximately 191 MB of active memory for text-only weights on the Google Pixel 11 Pro, while the full multimodal model requires approximately 567 MB.

EmbeddingGemma 2's score on the MTEB Code benchmark improved from 68.76 in the first generation to 78.68, an increase of 9.92 points.

EmbeddingGemma 2 is the first to unify text, code, images, video, and audio into a single embedding space, and is optimized for edge deployment to support local RAG.

EmbeddingGemma 2 employs Matryoshka Representation Learning technology, allowing embedding vectors to be truncated to 512, 256, or 128 dimensions without severely degrading representational quality.

EmbeddingGemma 2 is a multimodal embedding model built by Google on the Gemma 4 architecture, capable of encoding text, images, audio, and video into a shared 768-dimensional vector space.

In addition to text and images, EmbeddingGemma2 also supports audio and video, converting content of different modalities into 768-dimensional vectors.

SuperLocalMemory integrates Google's EmbeddingGemma 2 model and supports finding images and PDF files through natural language descriptions.

EmbeddingGemma 2 is a text embedding model in Google's Gemma series, specifically used to map text into dense vectors rather than generative conversation, and is suitable for tasks such as RAG, semantic retrieval, clustering and deduplication.

EmbeddingGemma 2's context window was expanded from v1's 2048 tokens to 8192, MTEB Code improved from 68.76 to 78.68, and it already supports Ollama, llama.cpp, sentence-transformers, MLX, vLLM, and browser-side transformers.js; there is currently no independent retrieval benchmark test for v2, and the SciFact score of 0.7861 belongs to v1.

Qwen3-Embedding, EmbeddingGemma 2, and nomic-embed-text v1.5 are all Apache 2.0 licensed, BGE-M3 is MIT licensed, and all can be used directly in commercial products; EmbeddingGemma v1 is governed by the Google Gemma terms; jina-embeddings-v5-text-small uses the CC BY-NC 4.0 license, and commercial use requires obtaining a separate license.

The per-text cosine similarity between the int8 quantized repack and the original float32 version ranges from 0.9888 to 0.9978; the cosine similarity between bf16 L4 and the original float32 reaches 0.9999 to 1.0000.

On the SciFact dataset (300 queries, 5,183 scientific abstracts, 768 dimensions), the int8+int4 repack model retained 97.9% of the original float32 version's nDCG@10 (0.836 vs 0.854).

EmbeddingGemma 2 comes with vision and audio encoders, but this project only loads the text tower.

EmbeddingGemma 2 is an open-source text embedding model that converts text into 768-dimensional vectors.

EmbeddingGemma 2 uses Matryoshka vector representations, where the first 128, 256 or 512 of the 768 dimensions can be used independently as embeddings.

EmbeddingGemma 2 very likely adopts a structure of a shared backbone plus modality adapters, rather than configuring a complete encoder for each modality.

EmbeddingGemma 2 has 740M parameters.

The currently public information about EmbeddingGemma 2 focuses on model specifications and benchmark scores, and its actual retrieval quality still needs to be verified on different datasets.

The heat and power consumption performance of multimodal inference on phones directly affects whether it can stay resident in an application.

The bottleneck of on-device RAG is memory: model weights plus runtime overhead must be kept within a range acceptable to the application, so the 191MB text mode is easier to integrate than the 567MB full mode.