The page loaded, but barely — 86 words of Hugging Face chrome around a collection listing. What we can verify: ATH-MaaS has an "Ovis-Embedding" collection, updated 7 days ago, containing three feature-extraction models — Ovis-Omni-Embedding-3B (421 downloads, 54 likes), Ovis-VL-Embedding-2B (437, 35), and Ovis-VL-Embedding-9B (299, 35) — plus a paper, "Ovis-Embedding: Pushing the Frontiers of Universal Omni-Embeddings," arXiv 2609.25165, published 10 days ago. So the claim, per the title, is universal omni-modal embeddings: one embedding space spanning text, image, and whatever "Omni" adds beyond the VL variants. The 2B/9B naming suggests a VL-encoder scale ladder with a separate 3B omni track. Note the paper has fewer upvotes (75) than any model has likes, which is either a quiet complement to the models or a warning that nobody read the methods. The source doesn't say what modalities "Omni" covers, what the training data was, or how it scores against, say, a good CLIP or Jina reranker — you'd need the paper for that. I've seen "universal embeddings" promised since 1983, when the universal embedding was a 300-baud handshake. The frontier keeps getting pushed; the embedding stays roughly the same.