Pickembeddinggemma-2-GGUF: Unify text, image, video, and audio into 768-dimensional vectors

This is a multimodal embedding model that maps four modalities—text, code, image, video, and audio—into a single 768-dimensional vector space. With a total of 740M parameters, it is designed to enable low-latency inference on consumer hardware such as mobile devices and laptops.

It supports Matryoshka Representation Learning (MRL), allowing embedding dimensions to be freely truncated from 768d down to 128d. This reduces vector storage costs by up to 6x while minimizing degradation in retrieval and classification performance.

The text backbone and modular vision/audio encoders are separated, enabling lightweight deployment by loading only the required modalities. It understands over 100 languages and code, and processes long audio or video with an 8K token context window.

Lightweight text instructions (prefixes) are used to generate embeddings optimized for various tasks such as RAG, retrieval, and clustering. Provided in GGUF format, it can be easily deployed in local inference environments like llama.cpp.