E5 Omni 7B

e5-omni-7B is an omni-modal embedding model built on Qwen2.5-Omni-7B that produces a unified embedding space for text, images, audio, and video, enabling cross-modal retrieval with a single model.

Key Strengths

Unlike models that rely on implicit alignment from vision-language backbones, e5-omni applies three explicit alignment techniques:

• modality-aware temperature calibration to align similarity scales across modalities
• a controllable negative curriculum with debiasing that focuses on confusing negatives while reducing false negative impact
• batch whitening with covariance regularization to match cross-modal geometry in the shared embedding space

These components address common issues in omni-modal embeddings: inconsistent score sharpness, imbalanced in-batch negative hardness, and mismatched first- and second-order statistics across modalities.

Benchmark Performance

The model achieves strong results on the MMEB-V2 and AudioCaps benchmarks.

Full experimental results and comparisons to bi-modal and omni-modal baselines are documented in the associated paper (arXiv:2601.03666). The explicit alignment recipe also transfers to other VLM backbones.