WeChat's 2B multimodal embedder tops prior 8B open models; 9B reaches 80.6 on MMEB-v2

WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report

Junjie Zhou, Ke Mei, Lei Li, Tianyi Wang, Fengyun Rao, Jing Lyu

cs.CV, cs.CL, cs.IR

2026-08-25

WeMM-Embedding adapts Qwen3.5 into 2B/4B/9B universal multimodal embedders. The 2B model scores 77.9 on MMEB-v2, above the previous leading 8B open baseline; the 9B model reaches 80.6 and is deployed in WeChat search and recommendation.

What problem this solves

Retrieval, recommendation, classification, and agents all want one vector space for text, images, videos, and documents. CLIP-style dual encoders do not naturally handle interleaved inputs, videos with transcripts, or composed queries. MLLMs can read those inputs, but they still need large-scale alignment and finer relevance training before they are useful embedders. Open 2B models such as Qwen3-VL-Embedding and DME already sit in the mid-70s on MMEB-v2, and scaling those recipes to 8B has not opened a large gap.

Method

All three sizes sit on native-multimodal Qwen3.5. An <embedding> token is appended, last-layer hidden state is pooled and L2-normalized. Several embedding tokens can share one sequence, so a video plus ASR can emit a video-only vector and a joint vector in one pass. Matryoshka training makes every prefix from 64-d up to full width a valid embedding.

Training examples share one pair format: optional instruction, source, positive target, optional hard negatives, optional graded score. Stage 1 trains on several hundred million pairs: weak web pairs, captions, retrieval, classification rewritten as source-label, multimodal QA, and human graded relevance. Each batch stays inside one task so in-batch negatives share a candidate space. Near-duplicates are masked to cut false negatives on shared labels. Graded batches use a score-gap-weighted CoSENT loss.

Stage 2 is about one tenth as large. Semantic IDs from three-level residual k-means down-sample frequent codes. An MLLM filters mismatched pairs and repairs dirty alt-text, then hard negatives are mined. Selected tasks add reranker-score ranking; the 2B and 4B models also distill bidirectional KL from a frozen 9B teacher. The 9B model has no larger teacher, so several Stage-2 specialists are merged. Reranker supervision is kept only where it actually helps.

Results

MMEB-v2, 78 datasets:

ModelSizeOverall
Qwen3-VL-Embedding2B73.2
DME-Small2B74.8
WeMM-Embedding2B77.9
WeMM-Embedding4B79.2
WeMM-Embedding9B80.6

The 2B model is 4.7 above same-size Qwen3-VL-Embedding, 3.1 above DME-Small, and slightly above the previous leading 8B open baseline. The 9B model led the official board on 2026-08-24. On MMEB-v3, audio tasks are scored zero because the model has no audio; V3-All is 56.0 / 58.2 / 59.5 for 2B / 4B / 9B, with the best Text and Agent group scores in the table. On the 12-dataset cross-modal suite from the Gemini Embedding 2 report, averages are 79.8 / 80.8 / 81.7.

On 26 in-house WeChat tasks the 2B model averages 72.0 against an open baseline at 60.9, winning all five categories. Fourteen online A/B tests were positive and the model is live in Channels, Official Accounts, Moments, and e-commerce. Per-test lifts are not reported.

MRL on the 2B model keeps 98.7% of 2048-d image and video scores at 256-d, about 99% at 512-d. Visual documents lose more. In a small Stage-1 ablation, dropping task-consistent batching cuts 71.9 to 68.5. Stage 2 walks 75.7 to 77.9, with distillation adding 0.9 and a larger visual budget another 0.3.

Why it matters

Universal multimodal embeddings are already production infrastructure. A 2B model beating older 8B recipes means one Qwen backbone can serve images, video, documents, and interleaved queries without a dual encoder per modality. 256-d vectors that keep most of the quality are the part that actually changes indexes.

This is a technical report. The transferable pieces are the pair format, task-consistent negatives, duplicate masks, and distillation for small models. The 26-task board and 14 A/B tests show deployment, not a reproducible effect size.

Limitations

No audio, so MMEB-v3 All is dragged by eleven zeroed tasks. The hundreds of millions of pairs are described at a high level, including licensing. In-house and online numbers have no public protocol and no reported A/B effect sizes. The 9B model is a merge of specialists, so data and merging are entangled. The leaderboard snapshot is 2026-08-24. The authors list omni-modal inputs and larger sizes as future work.

Terms

Source

Related papers

All paper explainers