Google released EmbeddingGemma 2, an open-weight multimodal embedding model based on Gemma 4 that maps text, code, images, video, and audio into a unified 768-dimensional vector space. Optimized for on-device deployment under the Apache 2.0 license, it features Matryoshka Representation Learning (MRL) to balance storage efficiency and retrieval quality.
Termsslug: matryoshka-representation-learning
Matryoshka Representation Learning
Matryoshka 表示学习
A technique allowing users to truncate embedding dimensions to trade off accuracy for computational efficiency within a single model.
Curated Intelligence
InfoQ 中文T2·Models
85 pts
Hugging Face BlogT1·Models
68 pts
IBM released the open-source Granite Embedding Multilingual R2 series based on ModernBERT, featuring 97M and 311M parameter models. The 97M variant achieves a state-of-the-art 60.3 score on MTEB Multilingual Retrieval for sub-100M models, while the 311M model scores 65.2, supporting Matryoshka embeddings and 32K context across 200+ languages.