[Models](/models)/[Embeddings](/models?category=embeddings)/tencent/WeMM-Embedding-2B

HFVisual EmbeddingsApache 2.0 (repo LICENSE; the HF card tags it "other")

# WeMM-Embedding-2B

by tencent

Tencent's 2B multimodal embedding model: text, image, video and visual documents in one 2,048-dimension space

424dl/month

2.72B total (Qwen3.5-2B backbone plus vision tower)params

[HuggingFace](https://huggingface.co/tencent/WeMM-Embedding-2B)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`tencent/WeMM-Embedding-2B`

## Deploy WeMM-Embedding-2B

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

WeMM-Embedding-2B is a universal multimodal embedding model from the WeChat team at Tencent, released 26 August 2026 on a Qwen3.5-2B backbone. Give it text, an image, a video, a scanned page, or several of those interleaved in one call, and it returns a single L2-normalized 2,048-dimension vector. Audio is the one input it refuses.

At 2.72B parameters this is the size most archives should try first. It posts a 77.9 MMEB-v2 average against 73.2 for Qwen3-VL-Embedding-2B and 59.3 for VLM2Vec-V2 at the same scale, and it beats every 8B model in the same table except Qwen3-VL-Embedding-8B, which it also passes. It even takes MMEB-v3 MCMR off its own 4B sibling, 42.5 against 41.9.

Matryoshka training lets you pick the output width when you build the index. The published truncation points are 64, 128, 256, 512, 1,024 and 2,048, and Tencent measured 256-dimension embeddings retaining 98.7% of full-dimension image and video performance on MMEB-v2. Storing 256 floats instead of 2,048 is roughly an 8x cut in bytes per vector, which is usually the line item that decides how much of an archive you can afford to index.

## Architecture

Qwen3.5-2B language backbone with a 1,024-dimension vision tower, 2,720,809,792 parameters in BF16. The size in the name refers to the language model, so the safetensors total runs higher once the vision encoder is counted. Position embeddings extend to 262,144 tokens. Output is a 2,048-dimension L2-normalized vector with Matryoshka truncation at 64, 128, 256, 512, 1,024 and 2,048. Tencent publishes serving paths for Transformers, Sentence Transformers, vLLM 0.27.0 with the pooling runner, and SGLang 0.5.9.

## Mixpeek SDK Integration

IngestRetrieve

// WeMM-Embedding has no managed Mixpeek extractor yet, so encode with the
// model and upsert the vectors. Route verified against the live OpenAPI:
// POST /v1/namespaces/{namespace\_id}/documents/upsert (document\_id + vectors
// are the only required fields).
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns\_your\_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API\_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection\_id: "col\_your\_collection",
      documents: \[
        {
          document\_id: "reel-00412",
          // 2048 floats from model.embedding(...). Slice and renormalize first
          // if you are storing a truncated Matryoshka dimension.
          vectors: { "multimodal-embedding": wemmVector },
          payload: { title: "Q3 launch reel", duration\_s: 94 },
          metadata: { source: "brand-archive" },
        },
      \],
    }),
  },
);

## Capabilities

*   One vector space covering text, image, video and visual-document queries
*   Matryoshka truncation to 64, 128, 256, 512, 1,024 and 2,048 without re-encoding
*   Interleaved inputs, so several images and a video can share a single embedding call
*   262,144-token position range for long visual documents
*   Chinese and English
*   Published vLLM 0.27.0 and SGLang 0.5.9 serving recipes

## Use Cases on Mixpeek

Text-to-video search over an archive, returning moments rather than filenames

Retrieving scanned pages and slide decks without an OCR stage in front of the index

Cross-modal dedup, where an image and its caption should land in the same neighbourhood

Shrinking an existing index to 256 dimensions once storage cost outgrows the accuracy it buys

## Benchmarks

Dataset

Metric

Score

Source

MMEB-v2 (78 datasets)

Average

77.9

WeMM-Embedding technical report, Table 1

MMEB-v2 Image

Hit@1

79.6

Technical report, Table 1

MMEB-v2 Video

Hit@1

70.8

Technical report, Table 1

MMEB-v2 VisDoc

NDCG@5

80.7

Technical report, Table 1

MMEB-v3 (190 tasks)

V3-All

56.0

Technical report, Table 2

MMEB-v3 Text

NDCG@5

45.3

Technical report, Table 2

MMEB-v3 Agent

Hit@1

45.1

Technical report, Table 2

MMEB-v3 MCMR

Hit@1

42.5

Beats the 4B's 41.9; Table 2

MMEB-v3 Audio

Hit@1

0.0

Audio is unsupported and scores zero by construction

## Performance

Input SizeText, image, video, visual document, or interleaved; 262,144-token position range

Embedding Dim2048 (Matryoshka down to 64)

GPU LatencyInput dependent

GPU ThroughputBatch dependent

GPU Memory~5 GB in BF16 for weights alone

Tencent publishes no latency or throughput figures for this model. The memory number is computed from the 2,720,809,792 BF16 parameters in the safetensors index and covers weights only, so serving needs headroom for activations and for batched video frames.

## Common Pipeline Companions

[

tencent/WeMM-Embedding-9B

Larger size, +2.7 MMEB-v2 average and +3.5 on video

](/model/tencent/WeMM-Embedding-9B)[

nvidia/omni-embed-nemotron-3b

Omni-modal alternative at similar scale, with audio

](/model/nvidia/omni-embed-nemotron-3b)[

Qwen/Qwen3-VL-4B-Instruct

Caption or answer over the retrieved candidates

](/model/Qwen/Qwen3-VL-4B-Instruct)

## Frequently Asked Questions

### Can I run WeMM-Embedding on Mixpeek today?

Not as a managed extractor. GET /v1/discovery/extractors returns 13 shipping extractors and none of them accepts an arbitrary Hugging Face model id, so a collection cannot be pointed at these weights. Two paths work now. Run the model on your own hardware and upsert the vectors through POST /v1/namespaces/{namespace\_id}/documents/upsert, which takes a named-vector map and stores it in Mixpeek Vector Store alongside everything else. Or package it as a custom extractor plugin, which is available on Enterprise namespaces.

### Does WeMM-Embedding handle audio?

No. The model card states that audio input is not supported, and on the 11 audio tasks in MMEB-v3 it scores 0.0, because the benchmark assigns zero to tasks a model cannot attempt. If audio has to live in the same index, the published omni-modal alternatives are Omni-Embed-Nemotron at 36.5 and E5-Omni-7B at 43.0 on the same tasks.

### What licence does WeMM-Embedding carry?

Apache 2.0. The Hugging Face card sets the license field to "other", which reads as restrictive, and the LICENSE file in the repo resolves it: Tencent states the model is "licensed under Apache-2.0" except for listed third-party components, which keep their original terms. No additional restrictions are imposed beyond those.

### Is the 2B good enough, or should I start with the 9B?

The 2B recovers 96.7% of the 9B's MMEB-v2 average, 77.9 against 80.6, on 29% of the parameters. Where the 9B pulls away is video (74.3 against 70.8) and the MMEB-v3 agent tasks (51.0 against 45.1). Index a sample with the 2B, score it on your own queries, and move up only if video recall is the thing that falls short.

### How small can the embedding go before quality drops?

Tencent measured 256-dimension embeddings retaining 98.7% of full-dimension image and video performance on MMEB-v2, which is the only truncation figure they publish. The trained cut points are 64, 128, 256, 512, 1,024 and 2,048. Slicing to an arbitrary width is not supported: use a dimension listed in model.config.matryoshka\_dimensions, and renormalize after slicing.

## Explore on Mixpeek

[

More Embeddings models

Compare alternatives in this category

](/models?category=embeddings)[

Best Multimodal Embedding Models

Hand-picked tools & platforms compared

](/curated-lists/best-multimodal-embedding-models)[

How CLIP, SigLIP & CLAP Work

Deep-dive technical guide

](/guides/contrastive-learning-how-clip-siglip-and-clap-work)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

FrameworkHF

Organizationtencent

FeatureVisual Embeddings

Output768-dim vector

Modalitiesvideo, image

RetrieverVector Search

Parameters2.72B total (Qwen3.5-2B backbone plus vision tower)

LicenseApache 2.0 (repo LICENSE; the HF card tags it "other")

Downloads/mo424

[

### Research Paper

WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report

arxiv.org](https://arxiv.org/abs/2608.24053)

### Build a pipeline with WeMM-Embedding-2B

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

### Alternative Models

[

openai/clip-vit-large-patch14

Visual Embeddings

](/model/openai/clip-vit-large-patch14)[

google/siglip-base-patch16-224

Visual Embeddings

](/model/google/siglip-base-patch16-224)[

google/siglip2-giant-opt-patch16-384

Visual Embeddings

](/model/google/siglip2-giant-opt-patch16-384)[

facebook/dinov2-large

Visual Embeddings

](/model/facebook/dinov2-large)

### Related in Embeddings

[

laion/clap-htsat-fused

Audio Embeddings

](/model/laion/clap-htsat-fused)[

facebook/pe-av-large

Audio Embeddings

](/model/facebook/pe-av-large)[

tsinghua-ee/WAVE-7B

Audio Embeddings

](/model/tsinghua-ee/WAVE-7B)[

facebook/encodec\_24khz

Audio Embeddings

](/model/facebook/encodec_24khz)