Ops-MM-embedding-v1-2B

Ops-MM-embedding-v1-2B is a dense, large-scale multimodal embedding model developed and open-sourced by the Alibaba Cloud OpenSearch-AI team, fine-tuned from Qwen2-VL.

Key Features

Unified Multimodal Embeddings

• Encodes text, images, text-image pairs, visual documents, and videos (by treating video frames as multiple image inputs) into a unified embedding space for cross-modal retrieval.

High Performance on MMEB

• Achieves SOTA results among models of similar scale on MMEB-V2 and MMEB-Image benchmark (until 2025-07-03).

Multilingual Capabilities

• The larger variant (Ops-MM-embedding-v1-7B) achieves SOTA performance among dense models on the ViDoRe-v2 benchmark, demonstrating strong cross-lingual generalization.

Training data

MMEB-train, CC-3M, colpali training set.

Performance

MMEB-V2

Model	Model Size (B)	Overall	Image-Overall	Video-Overall	Visdoc-Overall
seed-1.6-embedding	unknown	71.27	77.78	55.34	73.44
Ops-MM-embedding-v1-7B	8.29	67.61	72.72	53.76	70.34
Ops-MM-embedding-v1-2B	2.21	63.44	69.03	47.56	66.96
VLM2Vec-V2.0-Qwen2VL-2B	2.21	58.02	64.85	34.85	65.36
gme-Qwen2-VL-7B-Instruct	8.29	57.83	55.95	38.43	75.18
gme-Qwen2-VL-2B-Instruct	2.21	54.08	51.89	33.64	72.71

MMEB-Image

The table below compares performance on MMEB-Image benchmark among models of similar size.

Model	Model Size (B)	Image-Overall	I-CLS	I-QA	I-RET	I-VG
Ops-MM-embedding-v1-2B	2.21	69.03	68.07	65.11	69.17	80.85
B3_Qwen2_2B	2.21	68.1	67	61.19	70.85	79.88
LLaVE-2B	1.95	65.2	62.1	60.2	65.2	84.9

ViDoRe-v2

Model	Avg	ESG Restaurant Human	MIT Bio Multi.	Econ Macro Multi.	ESG Restaurant Synth. Multi.
gme-7B	55.61	63.37	49.49	54.21	55.38
seed 1.6 embedding	56.57	63.3	57.14	53.85	51.99
Ops-MM-embedding-v1-7B	59.59	66.27	54.34	60.92	56.82
Ops-MM-embedding-v1-2B	53.18	58.57	52.87	47.89	53.39

Usage

from ops_mm_embedding_v1 import OpsMMEmbeddingV1, fetch_image


model = OpsMMEmbeddingV1(
    "OpenSearch-AI/Ops-MM-embedding-v1-2B",
    device="cuda",
    attn_implementation="flash_attention_2"
)

t2i_prompt = "Find an image that matches the given text."
texts = [
    "The Tesla Cybertruck is a battery electric pickup truck built by Tesla, Inc. since 2023.",
    "Alibaba office.",
    "Alibaba office.",
]
images = [
    "https://upload.wikimedia.org/wikipedia/commons/e/e9/Tesla_Cybertruck_damaged_window.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/e/e0/TaobaoCity_Alibaba_Xixi_Park.jpg",
    "https://upload.wikimedia.org/wikipedia/commons/thumb/b/b0/Alibaba_Binjiang_Park.jpg/1024px-Alibaba_Binjiang_Park.jpg"
]

images = [fetch_image(image) for image in images]

# Text and image embedding
text_embeddings = model.get_text_embeddings(texts)
image_embeddings = model.get_image_embeddings(images)
print('Text and image embeddings', (text_embeddings @ image_embeddings.T).tolist())

# Fused Embedding
text_with_image_embeddings = model.get_fused_embeddings(texts=texts, images=images, instruction=t2i_prompt)
print('Text and image embeddings', (text_embeddings @ image_embeddings.T).tolist())

# Multi-image embeddings
multi_images = [
    [images[0]],
    [images[1], images[2]],
]
multi_image_embeddings = model.get_image_embeddings(multi_images)
print('Multi-image embeddings', (multi_image_embeddings @ multi_image_embeddings.T).tolist())