Tencent Releases WeMM-Embedding Family of Multimodal Embedding Models

Tencent released its general-purpose multimodal embedding family, WeMM-Embedding, on Hugging Face on August 25, followed by a technical report the next day. Built on Qwen3.5, the series comes in 2B, 4B, and 9B sizes, supports text, image, video, visual document, and arbitrary interleaved multimodal inputs, and uses MRL (Multi-Resolution Representation Learning) to return 4096-dimensional L2-normalized embedding vectors. Notably, as relayed by @reachsumit, even the 2B model surpasses an 8B baseline.

Confirmed

• The family includes 2B, 4B, and 9B versions built on the Qwen3.5 architecture
• Supports text, image, video, visual document, and interleaved multimodal inputs, outputting 4096-dim L2-normalized embeddings
• Achieves SOTA on the MMEB v2 and v3 benchmarks
• Per @NielsRogge's comparisons, the model leads on tasks such as COCO image-text retrieval (image-to-text)
• Training was two-stage: large-scale multimodal alignment followed by a stage using curated data
• Tencent says the model delivers state-of-the-art retrieval and recommendation performance on public benchmarks and in large-scale WeChat applications
• Audio input is not yet supported

Why it matters

A unified multimodal embedding space is a key component for cross-modal retrieval and recommendation; WeMM-Embedding maps text, images, video, and documents into a shared representation space and tops MMEB, with even its small versions proving competitive—making it highly practical for retrieval and recommendation applications.

2026-08-25 ~ 2026-08-26 · 5 related posts

• Tencent releases WeMM-Embedding-2B multimodal embedding model — tencent · 2026-08-25
• Tencent releases WeMM-Embedding multimodal model family — jacek2023 · 2026-08-25
• [source] Tencent releases universal multimodal embedding model, hits SOTA on MMEB — NielsRogge · 2026-08-25

2 near-duplicate retellings: tencent · _reachsumit