← All models

WeMM-Embedding-4B

Published Publisher Tencent

Positioning

WeMM-Embedding-4B is a universal multimodal embedding model developed by Tencent’s WeChat Vision team and built on Qwen3.5. It produces representations for retrieval and similarity tasks rather than generating conversational answers. (huggingface.co)

Capabilities and uses

  • Accepts text, images, videos, visual documents and interleaved multimodal inputs in a shared embedding space.
  • Returns a 2,560-dimensional, L2-normalized vector. Matryoshka embeddings support dimensions of 64, 128, 256, 512, 1,024 and 2,560; shorter vectors are obtained by truncation and re-normalization.
  • Intended uses include multimodal search, cross-modal retrieval, similarity matching and recommendation. (github.com)

Limitations and availability

  • Audio input is not supported. Reduced dimensions may trade retrieval quality for storage efficiency.
  • Tencent publishes model weights and code under Apache 2.0; third-party components retain their own licenses. Self-hosting requires suitable computing resources. (huggingface.co)

Date

The supplied technical report was submitted on August 25, 2026. This is the verified report submission date, not necessarily a separately verified model-weight release date. (arxiv.org)

Pricing

As of November 16, 2026, a model-specific official mainland China CNY price and international USD price could not be verified from the official pricing materials searched. Input and output API billing rates are therefore unverified; no amount or billing unit is assumed. Open weights do not make deployment infrastructure free. (cloud.tencent.com)

Official sources