最新模型
Google Launches EmbeddingGemma 2, a Modular Encoder for On-Device Multimodal Retrieval
On October 6, Google released EmbeddingGemma 2, which maps text, code, images, video, and audio into a shared vector space and makes its weights available under Apache 2.0. The model can load encoders as needed, but quantized memory figures and full-precision evaluation results should be interpreted separately.
Google DeepMind released EmbeddingGemma 2 on October 6, extending its previously text-focused embedding model to multimodal retrieval. Based on Gemma 4, the new model has 740 million parameters, and its weights are released under Apache 2.0. It lets text queries match directly against image, audio, or video content, making it suitable for local media search and retrieval-augmented generation (RAG). Official announcement
The architecture is modular: the text and code component has 270 million parameters, the vision encoder adds 170 million, and the audio encoder adds 300 million. After processing through a shared backbone, each modality is mapped into the same 768-dimensional space. In practice, a code-only index can omit the media encoders; cross-modal search may also simplify pipelines that first generate image captions or transcribe speech before creating text embeddings. The official developer guide provides integration instructions for Sentence Transformers 6.1.0 and later. Developer guide
Vector storage can also be adjusted. The model supports Matryoshka Representation Learning, allowing users to retain the first 512, 256, or 128 dimensions. After truncation, vectors must be L2-normalized again, and query and database vectors must use the same dimensionality. The model card reports a code evaluation score of 78.68 at 768 dimensions, falling to 71.41 at 128 dimensions. The sixfold reduction in vector capacity therefore comes with a quality tradeoff, and does not mean the entire vector database will shrink sixfold. Model card
Deployment figures should be distinguished from evaluation conditions. Google reports that, on a Pixel 11 Pro, the quantized version uses about 191 MB of active memory for the text weights, while the complete multimodal version uses about 567 MB. The model card’s benchmark results use full-precision checkpoints, so they should not be taken as evidence that quantized versions retain the same quality. Google also recommends BF16 or FP32 for standard floating-point inference, since FP16 may produce NaNs or silently reduce embedding quality. Memory figures, precision guidance
The weights are available from Hugging Face. Chinese technical teams should next measure recall, quantization error, and index capacity on Traditional Chinese documents, code, and mixed-media corpora; the official multilingual average score may not reflect performance for a particular language or domain. Model weights