Multimodal retrieval and unified representation are now core infrastructure for search, recommendation and agentic systems. WeMM-Embedding-9B provides a single 4,096-dimensional L2-normalized embedding space that encodes text, images, videos and visual documents — including arbitrarily interleaved multimodal inputs — so heterogeneous content can be compared directly in the same vector space. Built on a Qwen3.5 backbone and refined with a two-stage training curriculum, the 9B variant prioritizes both broad coverage and fine-grained relevance.
Key Capabilities
-
Unified multimodal encoding: encodes any subset or interleaving of text, images and videos into a single dense vector, enabling direct cross-modal similarity and retrieval without separate modality-specific pipelines. This simplifies pipelines for search, recommendation and multimodal RAG.
-
High-dimensional, Matryoshka-ready vectors: default embedding dimension is 4,096 with Matryoshka Representation Learning support (you can truncate to smaller dims and re-normalize), so you can trade off storage/latency against retrieval quality predictably.
-
Production-ready engineering and integrations: released with Transformers/SentenceTransformers compatibility, processor utilities for vision/video, and serving notes for vLLM/SGLang; weights and code are publicly available under Apache‑2.0, facilitating offline and online deployment.
-
Strong empirical performance: the family (2B/4B/9B) reports leading scores on multimodal benchmarks (WeMM paper/MMEB), with the 9B variant achieving top-line retrieval metrics in the paper’s evaluations, demonstrating consistent gains on image, video and visual-document tasks.
Who it’s for, and tradeoffs
Great fit if you need a single embedding space to power multimodal retrieval, cross-modal ranking, or unified feature storage for downstream ML systems — particularly when you can accommodate a large, bfloat16-optimized model and want SentenceTransformers-style APIs. Look elsewhere if you need audio embedding support (audio is explicitly unsupported), extremely low-memory on-device deployment, or a much smaller model footprint — the 9B weights expect server/GPU inference and BF16/FP16 runtime considerations.
Where it fits
Use WeMM-Embedding-9B when you want high-quality, production-capable multimodal embeddings that integrate easily with existing retrieval stacks (FAISS, SentenceTransformers, vLLM) and where cross-modal ranking accuracy matters. For lighter-weight or audio-capable use cases, consider smaller embedding families or specialized audio/text models.
