Retrieval and recommendation run on embedding models — every video and article your feed serves you is a vector distance computation under the hood. While the spotlight sits on chat and generation, embeddings are the adapter that wires multimodal content into large systems. Tencent's WeChat Vision team has now open-sourced its production multimodal embedding family, WeMM-Embedding: three sizes (2B, 4B, 9B) under Apache 2.0, with weights on Hugging Face and code plus evaluation on GitHub (arXiv: 2608.24053).

The Span in One Table

Per the official technical report, WeMM-Embedding supports text, images, videos, visual documents, and arbitrarily interleaved multimodal inputs, mapped into one shared vector space; embeddings are taken from the last-layer hidden state at a dedicated token and L2-normalized. Training runs in two stages: large-scale multimodal alignment first, then refinement with curated data, fine-grained relevance supervision, and cross-scale knowledge transfer.

Official MMEB-v2 results (78 datasets): the 2B model scores 77.9 overall, surpassing the previously leading 8B open-source baseline — Qwen3-VL-Embedding 8B sits at 77.8, VLM2Vec 8B at 53.2. The 9B model reaches 80.6 overall (image 81.9, video 74.3, visual document 83.3), the highest score in the table; the two closed-source submissions without public weights (DME-Small/Medium) do not exceed it either. These are vendor-reported benchmark numbers, and the reproduction code ships with the repository.

Matryoshka Dimensions and Serving

Two engineering details stand out. First, Matryoshka representations: the three sizes support variable output dimensions from 64 up to 2048/2560/4096, and the 2B model retains 98.7% of its full-dimensional image and video performance when truncated to 256 dimensions — vector-store storage costs can drop by an order of magnitude with almost no retrieval-quality loss. Second, the inference stack: the team validated vLLM 0.27.0 and SGLang 0.5.9 serving, recommends transformers 5.2.0, and samples video at 64 frames in its evaluation pipeline.

The Harsher MMEB-v3, and a Self-Reported Weakness

The team also published MMEB-v3 results: 190 tasks, including the 78 v2 tasks, 53 text tasks, 47 agent tasks, 11 audio tasks, and MCMR, with unsupported tasks scored as zero. WeMM 2B/4B/9B score 56.0, 58.2, and 59.5 respectively, above Qwen3-VL-Embedding in the same table (2B 50.9, 8B 53.5), and above E5-Omni and Omni-Embed-Nemotron as well. Notably, the audio column reads all zeros for WeMM — audio input is not currently supported — and the team left that weakness in the table rather than hiding it.

From Benchmarks to Moments

The real weight is online validation. The report claims substantial gains on a 26-task in-house benchmark and consistent improvements across 14 online A/B tests, with deployment across WeChat Channels, Official Accounts, Moments, and e-commerce recommendation and search. This is not a leaderboard-only academic model; it is an industrial component that has run the full pipeline inside WeChat's core surfaces.

So What

Embedding models are the plumbing of the RAG and agent era: multimodal understanding is worthless if the retrieval layer cannot carry it to users. Open-sourcing a production-grade universal multimodal embedder — weights and evaluation included — lays a foundation for multimodal retrieval ecosystems. And that glaring zero in the MMEB-v3 audio column is a reminder: "universal" always has one modality missing.

References: arXiv 2608.24053 (huggingface.co/papers/2608.24053) and github.com/Tencent/WeMM-Embedding.