AI
Omni-Embed-Mini: Binding Modalities Without Forgetting via Dense Distillation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 30/09/2026, 20:00:00 UTC
Briefing created: 02/10/2026, 11:15:43 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 09:18:25 UTC.
Briefings
Omni-Embed-Mini is a 0.9B-parameter model that maps text, speech, audio, images, video, and visually-rich documents into a single shared cosine space without updating text-side parameters. It achieves 49.57 nDCG@10 on MTEB-v2 BEIR-8 while being 2.7x to 9.5x smaller than open omni embedders. The 2.3B variant is competitive with Gemini-embedding-2 on the overall-modality average.
Key points
- The model maps six modalities (text, speech, audio, images, video, and visually-rich documents) into a single shared cosine space without updating text-side parameters.
- Training uses a Matryoshka SigLIP contrastive loss combined with an online hybrid hard-negative miner to sharpen negatives as the encoder improves.
Context and limitations
Coverage is limited to the six modalities tested; the 2.3B variant's performance on the overall-modality average is competitive with Gemini-embedding-2 but specific comparisons to closed models are not detailed.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 11:15:43 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifz672igwn7y2aqnyzjh7ctjqeztc3z6ewup2xhlahip237rsb4lu