Rankings
Every model on the leaderboard, scored as the mean of the modality benchmarks it is actually evaluated on. Read it with the coverage column: a specialist averaging one benchmark is not doing the same job as a system averaging six, and the means are not on a common scale. The scope control opens on full-coverage systems, which is the only like-for-like reading of this board.
All numbers are taken from Omni-Embed-Mini (EMNLP 2026 Findings), Figure 7 and Appendix Tables 3 to 8. embeddinggemma-300m and bge-large-en are text retrievers scored on the OCR'd page text alone, whereas every other model on the visual-doc board sees the page image, so those two bars are not a like-for-like comparison. gemini-embedding-2 is closed and its parameter count is undisclosed; it is shown for reference only.
e5-omni-7B
8.9B parameters
Omni-Embed-Mini-2.3B
2.3B parameters
gemini-embedding-2
? parameters
LCO-Embedding-Omni-7B
8.9B parameters
e5-omni-3B
4.7B parameters
omni-embed-nemotron-3B
4.7B parameters
LCO-Embedding-Omni-3B
4.7B parameters
BidirLM-Omni-2.5B
2.5B parameters
Omni-Embed-Mini-0.9B
935M parameters
Only systems that bind all six modalities into one embedding space. This is the like-for-like comparison, and the block the paper itself reports against.
| # | Model | Params | Coverage | Score | Text | Speech | Audio | Image | Video | Doc |
|---|---|---|---|---|---|---|---|---|---|---|
e5-omni-7B Omni embedders | 8.9B | 6/6 | 52.93 | 45.34 | 54.86 | 43.26 | 68.36 | 44.73 | 61.00 | |
Omni-Embed-Mini-2.3B Ours · Omni embedders | 2.3B | 6/6 | 51.39 | 47.94 | 48.86 | 33.44 | 64.80 | 55.18 | 58.10 | |
gemini-embedding-2 Closed API embedders | ? | 6/6 | 49.51 | 56.24 | 51.54 | 36.80 | 47.27 | 48.14 | 57.09 | |
| 4 | LCO-Embedding-Omni-7B Omni embedders | 8.9B | 6/6 | 49.34 | 42.54 | 54.61 | 45.91 | 54.92 | 43.34 | 54.72 |
| 5 | e5-omni-3B Omni embedders | 4.7B | 6/6 | 49.28 | 47.80 | 52.04 | 40.34 | 63.90 | 32.92 | 58.66 |
| 6 | omni-embed-nemotron-3B Omni embedders | 4.7B | 6/6 | 48.34 | 50.48 | 52.85 | 44.00 | 48.16 | 34.90 | 59.63 |
| 7 | LCO-Embedding-Omni-3B Omni embedders | 4.7B | 6/6 | 47.52 | 39.42 | 53.02 | 43.88 | 52.02 | 43.86 | 52.91 |
| 8 | BidirLM-Omni-2.5B Omni embedders | 2.5B | 6/6 | 47.28 | 48.56 | 55.80 | 39.03 | 45.98 | 41.23 | 53.06 |
| 9 | Omni-Embed-Mini-0.9B Ours · Omni embedders | 935M | 6/6 | 36.33 | 49.57 | 43.28 | 33.43 | 26.29 | 18.48 | 46.92 |
Scores are per-benchmark averages and are not comparable across tabs; each tab uses its own metric, shown above the table. Parameter counts are taken from each model’s published card; closed-source models with undisclosed sizes are marked “?”. Coverage counts the modalities this leaderboard scores a model on, taken from the paper's benchmark tables. It measures evaluated breadth, not everything an architecture could in principle encode. The board opens scoped to systems covering all six modalities; widen it with the scope control above.