ByteDance’s Douyin multimodal embedding model reaches SOTA on MMEB-v2

_reachsumit · x · 2026-08-04

ByteDance’s Douyin multimodal embedding model uses latent reasoning to hit SOTA on MMEB-v2

The technical report describes a two-stage multimodal embedding system for large-scale search and recommendation.

Why it matters

Real-world platforms such as Douyin, Xiaohongshu, and YouTube need embeddings that are both efficient at billion-scale indexing and fine-grained enough for hard matching.
The authors argue that contrastive models are efficient but too coarse, while CoT-style embedding methods are more discriminative but too expensive to serve online.

Training design

Stage 1: large-scale contrastive pretraining to build a unified multimodal embedding space.
Stage 2: adds semantic sufficiency via two training-only mechanisms:
Evidence-Grounded Typed Latent Reasoning
Cross-Conditional Reconstruction

Claimed outcome

The model achieves state-of-the-art results on MMEB-v2.
The extra latency on the query side is described as minimal because the reasoning components are used only during training.

More from Multimodal