WeMM-Embedding

WeMM-Embedding-2B is a universal multimodal embedding model built on Qwen3.5. It accepts text, images, videos, visual documents, and interleaved multimodal inputs, and returns a 2,048-dimensional L2-normalized embedding. Audio input is not supported.

Installation

pip install torch transformers==5.2.0 "qwen-vl-utils[decord]==0.0.14" \
  "sentence-transformers>=5.7.0" "accelerate>=1.1.0"

Transformers

import torch
from qwen_vl_utils import process_vision_info
from transformers import AutoModel, AutoProcessor

model_id = "tencent/WeMM-Embedding-2B"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_id, trust_remote_code=True, dtype=torch.bfloat16
).cuda().eval()

messages = [{"role": "user", "content": [
    {"type": "image", "image": "/path/to/image.jpg"},
    {"type": "video", "video": "/path/to/video.mp4"},
    {"type": "text", "text": "This can be any text input."},
]}]
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=False
)
images, videos, video_kwargs = process_vision_info(
    messages,
    image_patch_size=16,
    return_video_kwargs=True,
    return_video_metadata=True,
)
if videos is not None:
    videos, video_metadata = zip(*videos)
    videos, video_metadata = list(videos), list(video_metadata)
else:
    video_metadata = None
inputs = processor(
    text=text,
    images=images,
    videos=videos,
    video_metadata=video_metadata,
    return_tensors="pt",
    **video_kwargs,
).to("cuda")

with torch.inference_mode():
    embedding = model.embedding(**inputs)

Use any subset of the content items to encode text, image, or video independently.

Sentence Transformers

from sentence_transformers import SentenceTransformer

model_id = "tencent/WeMM-Embedding-2B"
model = SentenceTransformer(model_id, trust_remote_code=True)

queries = [
    "Which Llama 4 model variants are available?",
    "How is mapo tofu prepared?",
]
documents = [
    "Mapo tofu is a Sichuan dish of soft tofu simmered in a spicy, numbing sauce of chili bean paste and Sichuan peppercorn.",
    {
        "image": "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/llama4_hgf.png",
        "text": "Represent this image.",
    },
    {
        "video": "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/mapo_tofu.mp4",
        "text": "Represent this video.",
    },
]

query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# (2, 2048) (3, 2048)

similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[0.0683, 0.4972, 0.0309],
#         [0.7829, 0.1428, 0.5492]])

Each input is a string, a URL or path, a PIL.Image, or a dict combining image, video, and text keys. Put image or video before text so the prompt matches the ordering used above. Chat messages such as {"role": "user", "content": [{"type": "image", "image": ...}, {"type": "text", "text": ...}]} are also accepted, which is the way to interleave several images or videos in one input.

Matryoshka Embeddings

embedding_256 = torch.nn.functional.normalize(embedding[..., :256], dim=-1)

With Sentence Transformers, pass truncate_dim and let it renormalize:

embeddings_256 = model.encode_document(documents, truncate_dim=256, normalize_embeddings=True)

Use a dimension listed in model.config.matryoshka_dimensions. On MMEB-v2, 256-dimensional embeddings retain 98.7% of the full-dimensional image and video performance.

Serving

vLLM 0.27.0:

MODEL_PATH=/path/to/WeMM-Embedding-2B
vllm serve "$MODEL_PATH" \
  --runner pooling \
  --chat-template "$MODEL_PATH/embedding_chat_template.jinja"

SGLang 0.5.9:

MODEL_PATH=/path/to/WeMM-Embedding-2B
python patch_sglang_video.py
python -m sglang.launch_server \
  --model-path "$MODEL_PATH" \
  --is-embedding \
  --enable-precise-embedding-interpolation

Evaluation

MMEB-v2

Results on 78 datasets from Table 1 of the technical report. Image and video tasks use Hit@1, while visual-document tasks use NDCG@5. Higher is better.

† Closed-source leaderboard submission without publicly released model weights or a public inference endpoint.

MMEB-v3

Results on all 190 tasks from Table 2 of the technical report. V3-All includes the 78 MMEB-v2 tasks, 53 text tasks, 47 agent tasks, 11 audio tasks, and MCMR. Unsupported tasks are assigned a score of zero.