GeeVec-Embeddings-mm-1.0

GeeVec-Embeddings-mm-1.0 is a multimodal embedding model, with approximately 9.41B parameters, built on top of Qwen3.5-9B. It is optimized for multimodal retrieval tasks and maps different modalities into a shared embedding space:

• text: text queries and passages for semantic retrieval.
• image: images and image-text inputs for cross-modal retrieval.
• video: video content represented by sampled frames.
• visual document: document-page images, including charts, tables, and slides.

Introduction

This repository hosts the model GeeVec-Embeddings-mm-1.0.

Technical highlights:

• Model Type: Multimodal Embedding
• Backbone: Qwen3.5-9B
• Total Parameters: approximately 9.41B
• Sequence Length: 8,192 tokens in the current embedding pipeline
• Embedding Dimension: 4096
• Supported Modalities: text, image, video, and visual document
• Pooling Method: last-token pooling
• Similarity: cosine similarity

Model API usage documentation: https://www.geevec.com/documentation. & https://geeknow.geevec.com/embedding_model/

Usage With GeeVec-Embeddings-mm-1.0

Using Hugging Face Transformers

Install Transformers with Qwen3.5 support and the public vision utilities:

pip install torch "transformers==5.8.0" accelerate qwen-vl-utils pillow decord

import os
import torch
from modeling import GeeVecEmbedding

model_path = os.environ["MODEL_PATH"]
model = GeeVecEmbedding.from_pretrained(
    model_path,
    dtype=torch.bfloat16,
    device_map="cuda:0",
)


queries = [
    {
        "text": "A person riding a bicycle along a city street.",
        "instruction": "Find an image that matches the given description.",
    },
    {
        "text": "A beach with clear water and white sand.",
        "instruction": "Find an image that matches the given description.",
    },
]
images = [
    {"image": "/path/to/bicycle.jpg"},
    {"image": "/path/to/beach.jpg"},
]

with torch.inference_mode():
    query_embeddings = model.encode(queries)
    image_embeddings = model.encode(images)

similarity = query_embeddings @ image_embeddings.T
print(similarity.cpu().tolist())

Visual-Document Retrieval

queries = [
    {
        "text": "Which page contains the company's annual revenue?",
        "instruction": "Retrieve a document page that answers the question.",
    },
]
pages = [
    {"image": "/path/to/document_page_1.png"},
    {"image": "/path/to/document_page_2.png"},
]

with torch.inference_mode():
    query_embeddings = model.encode(queries)
    page_embeddings = model.encode(pages)

similarity = query_embeddings @ page_embeddings.T
print(similarity.cpu().tolist())

Text-to-Video Retrieval

queries = [
    {
        "text": "A person preparing a meal in a kitchen.",
        "instruction": "Retrieve a video that matches the given description.",
    },
]
videos = [
    {
        "video": "/path/to/video.mp4",
        "fps": 1.0,
    },
]

with torch.inference_mode():
    query_embeddings = model.encode(queries)
    video_embeddings = model.encode(videos)

similarity = query_embeddings @ video_embeddings.T
print(similarity.cpu().tolist())

Notes

• The model produces 4096-dimensional embeddings for all supported modalities.
• Use task-specific instructions for queries and follow the corresponding evaluation configuration when reproducing benchmark results.
• Document pages are provided as images; convert PDF files to page images before encoding.
• Video paths are decoded and sampled by qwen-vl-utils; fps controls the requested sampling rate. Pre-sampled frame lists are also supported.

Evaluation

MMEB-v2 Retrieval

Results on the 44 retrieval datasets of MMEB-v2: 12 image retrieval, 5 video retrieval, 3 moment retrieval, and 24 visual-document retrieval tasks. Image and video tasks use Hit@1, while visual-document tasks use NDCG@5 (linear gains). Higher is better. Scores are percentages.

Video includes moment retrieval; AVG averages the 44 retrieval datasets equally.