LCO Embedding Omni 3B

LCO-Embedding/LCO-Embedding-Omni-3B is a multimodal embedding model that accepts text, image, audio, and video inputs and produces a unified vector representation. It is built on the Qwen2.5 Omni thinker component (dropping the talker) and implements the language-centric omnimodal representation learning framework described in the NeurIPS 2025 paper "Scaling Language-Centric Omnimodal Representation Learning."

A key strength is its state-of-the-art performance on the MIEB (Massive Image Embedding Benchmark), which covers 130 tasks across 8 categories and 38 languages. The model also supports cross-modal and combined-modality queries—for example, embedding a document that pairs text with a video or an image with audio—enabling retrieval across diverse media types.

The model uses a simple, fixed prompt template ("Summarize the above <modality> in one word:") baked into its chat template. Example retrieval similarities from the model card illustrate its capability: text-to-text similarity of 0.6199 (matching Mount Everest description), image-to-text similarity of 0.4396, audio-to-text similarity of 0.3809, and video-to-text similarity of 0.6406.

Underlying the approach is the Generation-Representation Scaling Law (GRSL), which posits that representational quality from contrastive fine-tuning scales positively with the underlying MLLM’s generative capabilities. The model is hosted as a managed, OpenAI-compatible API on gigarouter—no local installation or hardware configuration required.

Model Variants: The 3B variant described here; a 7B variant (LCO-Embedding-Omni-7B) is also available separately.

Key Reference: arXiv:2510.11693