Omni-Embed-Mini Rivals Gemini Embedding 2 at 2.3B Params

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, and it does so without updating any text-side parameter. The headline retrieval score is 49.57 nDCG@10 on MTEB-v2 BEIR-8.

The mechanism, described in the arXiv paper accepted to Findings of EMNLP 2026, is that the teacher signal needs no separate embedding model: each media sample is paired with a dense cascaded caption, and the target is simply the frozen backbone's own embedding of that caption. "Because teacher and student share the same backbone weights, they inhabit byte-identical geometry," the paper reports, so lightweight projectors plus phased LoRA adapters on the modality encoders carry the alignment. Training combines a Matryoshka SigLIP contrastive loss with an online hybrid hard-negative miner whose negatives sharpen as the encoder improves.

Because the text weights stay bit-identical to the backbone, the authors argue training "cannot regress text retrieval" while adding five more modalities, and they claim the 0.9B model is "~2.7x to 9.5x smaller than every open omni embedder we compare against." A 2.3B variant swaps in a native vision-language backbone and, in the authors' words, "is competitive with the closed gemini-embedding-2, edging ahead of it on the overall-modality average."

The abstract does not list the specific open embedders in the size comparison, nor publish per-modality breakdowns. It lands amid a steady run of smaller-footprint open releases we've tracked this week on the open-source feed.