LCO Embedding Omni 7B

LCO-Embedding-Omni-7B is a language-centric omnimodal embedding model that generates unified vector representations for text, image, audio, and video inputs, built on the Qwen2.5 Omni thinker component.

Key Strengths

• Omnimodal support: Processes and embeds text, images, audio, and video — including multimodal documents that combine multiple modalities (e.g., a video with text description).
• State-of-the-art on MIEB: Achieves top performance on the Massive Image Embedding Benchmark (MIEB), which covers 130 tasks across 8 categories (retrieval, document understanding, visual STS, zero-shot classification, few-shot linear probing, clustering, compositionality evaluation, and vision-centric QA) in 38 languages.
• Generation-Representation Scaling Law (GRSL): The paper establishes a formal link between a model's generative capabilities and its representation upper bound. GRSL shows that improving generative pretraining raises the ceiling for contrastive refinement, and is validated on the new SeaDoc benchmark for visual document retrieval in Southeast Asian languages.
• Peer-reviewed research: The work is accepted at NeurIPS 2025 and developed by DAMO Academy, Alibaba Group.

Architecture

The model uses only the thinker component of Qwen2.5 Omni. A baked-in instruction ("Summarize the above <modality> in one word:") is applied via the chat template, enabling consistent embedding behavior across all input types without requiring user-provided prompts.

Benchmark Performance

LCO-Embedding-Omni-7B sets a new state-of-the-art on MIEB, outperforming prior multimodal embedding models across diverse tasks. Specific scores are available in the paper and MIEB leaderboard.

For detailed evaluation and comparisons, refer to the NeurIPS 2025 paper and the model collection.