AntEmbed runs two retrieval modes on one index: its 8B generative embeddings score 81.5 on MMEB-v2
Synopsis
Multimodal retrieval previously forced a choice: discriminative embeddings encode fast but miss fine-grained evidence, while generative embeddings reason before encoding and carry richer semantics at extra latency. AntEmbed treats the original and reasoning-enhanced forms of a candidate as equivalent views of one instance, aligns them in a unified space with a single embedding token and multi-positive contrastive learning, then distills the reasoning branch's fine-grained information into the original branch. On MMEB-v2, AntEmbed-8B reaches an overall 81.5 with generative embeddings, 3.1 points above the strongest baseline DME-9B; its discriminative embeddings need no extra reasoning, and AntEmbed-2B scores 70.2 in base mode.
Interpretation
AntEmbed lets one multimodal retriever offer both fast encoding and reasoning-enhanced embeddings while sharing a single candidate index. UME-R1 previously modeled the two modes with two separate embedding tokens; the same instance showed low similarity across modes, so a retrieval system had to fix the mode in advance and keep separate corpus representations per mode. On MMEB-v2, AntEmbed-8B with generative embeddings scores 81.5 overall, 3.1 points above the strongest baseline DME-9B at 78.4 and 3.7 above Qwen3-VL-Embedding-8B at 77.8; its image, video and visual-document scores are 83.9, 74.8 and 82.7.
Training proceeds in two steps: original and reasoning-enhanced inputs are first matched inside one embedding space, then the reasoning branch's fine-grained information is distilled into the original branch. Multi-positive contrastive learning places both the original and the reasoning-enhanced candidate in the positive set for each query and requires every positive to outrank the negatives, rather than only raising the total probability of the positive set. In the ablation, first-stage multi-positive training alone averages 69.1, and adding second-stage cross-mode self-distillation raises it to 70.2; either similarity-distribution distillation or query representation alignment alone beats the first stage, and enabling both is best.
The reasoning content is generated by an external reasoner around query intent and the evidence that separates confusable candidates. A candidate's reasoning-enhanced form is not a description of itself in isolation: Qwen3-VL-Embedding-8B first retrieves the five most similar candidates as a confusion set, and the reasoner then names the attributes, relations and visual or textual evidence that distinguish the candidate from those neighbors. Swapping reasoners, a fine-tuned Qwen3-VL-8B-Instruct lifts the average from 73.3 in base mode to 78.1, Qwen3-VL-30B-A3B-Instruct reaches 79.3, and Gemini-3.1-Pro-Preview reaches 80.2.
Perspective
Evaluation centers on the image, video and visual-document retrieval tasks of MMEB-v2, scored with Hit@1 for image and video and NDCG@5 for visual documents; in the MMEB-v3 extension, AntEmbed does not support audio input, so the 11 audio tasks count as zero in the V3-All score.
Generative embeddings still require an extra autoregressive reasoning pass, with higher latency and compute than discriminative encoding; a clear score gap remains between base mode and CoT mode, 70.2 versus 77.2 for AntEmbed-2B. The MMEB-v3 V3-All score is affected by counting audio tasks as zero. Future work plans to integrate AntEmbed into agent systems and use agent-generated reasoning traces and task context to improve retrieval, with no results yet.
