Is WeMM-Embedding free to use commercially?

Yes. Tencent releases the code it wrote for WeMM-Embedding under the Apache License 2.0, which permits commercial use, modification and redistribution. The repository notes that third-party components keep their original licenses, so check those if you reuse the full stack. The model weights for all three sizes are published on Hugging Face under tencent/WeMM-Embedding-2B, -4B and -9B.

How does WeMM-Embedding compare to Qwen3-VL-Embedding?

On the MMEB-v2 average reported in the WeMM-Embedding repository, the 9B model scores 80.6 against 77.8 for Qwen3-VL-Embedding 8B. At the small end the gap is wider: WeMM-Embedding 2B scores 77.9 versus 73.2 for Qwen3-VL-Embedding 2B. The 9B model also leads on the video split, 74.3 against 67.1.

Which WeMM-Embedding size should you pick?

The 2B version is the value pick. WeMM-Embedding 2B posts a 77.9 MMEB-v2 average, which edges past the 77.8 that Qwen3-VL-Embedding scores at 8B, so a quarter of the parameters buys roughly the same quality as last generation's large model. Move up to 4B (79.2) or 9B (80.6) when retrieval accuracy matters more than serving cost.

Can WeMM-Embedding output smaller vectors to save storage?

Yes. Every WeMM-Embedding size is trained with Matryoshka representation learning, so a long vector can be truncated to a shorter one without re-encoding. The 2B model spans 64 to 2048 dimensions, the 4B model 64 to 2560, and the 9B model 64 to 4096. Picking a smaller dimension shrinks your index at a graceful cost in accuracy.

Tencent · 2026-08-25 · major

WeMM-Embedding — Tencent's multimodal retrieval models top MMEB-v2

WeMM-Embedding is a family of open multimodal embedding models from Tencent's WeChat Vision Team, released in 2B, 4B and 9B sizes under Apache-2.0. The 9B model scores 80.6 average on MMEB-v2.

Hugging Face model card banner for Tencent WeMM-Embedding-9B

Tencent open-sourced WeMM-Embedding, three multimodal embedding models that turn text, images, video and documents into one vector space.

Quick facts

Maker	WeChat Vision Team, Tencent
Sizes	2B, 4B, 9B
License	Apache-2.0
Inputs	Text, images, video, visual documents, interleaved
Output dimensions	Matryoshka, 64 to 4096
Serving	Transformers, SentenceTransformers, vLLM, SGLang
Weights	Hugging Face, tencent/WeMM-Embedding-2B / -4B / -9B

Benchmarks

MMEB-v2 (average, 8B-9B class)

WeMM-Embedding 9B		80.6%
DME-Medium 9B		78.4%
Qwen3-VL-Embedding 8B		77.8%
GME 8B		59.2%
VLM2Vec 8B		53.2%

source ↗

What is it?

WeMM-Embedding adds a universal embedding family that handles text, images, video, visual documents, and mixed inputs where all of those are interleaved in one query. Tencent's WeChat Vision Team released three sizes — 2B, 4B and 9B — with weights and code on GitHub and Hugging Face under Apache-2.0. Every size uses Matryoshka output dimensions, so one model can emit a short vector or a long one.

How does it work?

A two-stage training recipe teaches the WeMM-Embedding models to place different media types in a shared vector space, so a text query can retrieve a video clip or a page scan without a separate model per modality. Matryoshka representation learning nests smaller vectors inside larger ones, which lets you truncate an embedding to 64 dimensions and still keep usable accuracy.

Why does it matter?

Teams building search over mixed content — chat logs, screenshots, slide decks, product video — usually stitch together one embedding model per media type and then fight the mismatch between vector spaces. One Apache-2.0 model covering all of them removes that plumbing, and the 2B version already matches 8B baselines, so the cost of running it is small.

Who is it for?

search and RAG engineers

Frequently asked questions

Is WeMM-Embedding free to use commercially?: Yes. Tencent releases the code it wrote for WeMM-Embedding under the Apache License 2.0, which permits commercial use, modification and redistribution. The repository notes that third-party components keep their original licenses, so check those if you reuse the full stack. The model weights for all three sizes are published on Hugging Face under tencent/WeMM-Embedding-2B, -4B and -9B.
How does WeMM-Embedding compare to Qwen3-VL-Embedding?: On the MMEB-v2 average reported in the WeMM-Embedding repository, the 9B model scores 80.6 against 77.8 for Qwen3-VL-Embedding 8B. At the small end the gap is wider: WeMM-Embedding 2B scores 77.9 versus 73.2 for Qwen3-VL-Embedding 2B. The 9B model also leads on the video split, 74.3 against 67.1.
Which WeMM-Embedding size should you pick?: The 2B version is the value pick. WeMM-Embedding 2B posts a 77.9 MMEB-v2 average, which edges past the 77.8 that Qwen3-VL-Embedding scores at 8B, so a quarter of the parameters buys roughly the same quality as last generation's large model. Move up to 4B (79.2) or 9B (80.6) when retrieval accuracy matters more than serving cost.
Can WeMM-Embedding output smaller vectors to save storage?: Yes. Every WeMM-Embedding size is trained with Matryoshka representation learning, so a long vector can be truncated to a shorter one without re-encoding. The 2B model spans 64 to 2048 dimensions, the 4B model 64 to 2560, and the 9B model 64 to 4096. Picking a smaller dimension shrinks your index at a graceful cost in accuracy.

Try it

pip install sentence-transformers, then load tencent/WeMM-Embedding-2B