Eddy-VL Embedding v2

Urock-AI · urock.kr · License: Apache 2.0

Eddy-VL Embedding v2 is a multimodal embedding model light enough to run on edge devices. It keeps most of the retrieval quality of WeMM-Embedding-2B in a lighter package — built at Urock-AI Lab for real-world multimodal search over images, video, and documents.

At a glance

Lighter than WeMM-Embedding-2B. Three linear-attention layers are removed from the Gated DeltaNet stack. embedding() uses 2.036B parameters instead of 2.213B. On long text, compute is 12% lower. On a max-size image, it is 8.4% lower.
Close to WeMM-Embedding-2B on the same evaluation. Hit@1 reaches 99.7% of WeMM-Embedding-2B on SC, 97.7% on images, 98.5% on visual documents, and 97.4% on video.
One embedding space for images, video, and documents. The vector is the L2-normalized hidden state at the final <embedding> token.

BF16 multimodal embedding model. Three linear-attention layers are removed from the Gated DeltaNet stack of tencent/WeMM-Embedding-2B, and the remaining stack is restored by distillation from that model. Softmax attention layers are left in place. WeMM-Embedding-2B is built on Qwen/Qwen3.5-2B, and this model keeps the same hybrid backbone. embedding() returns one L2-normalized vector, taken from the last non-padding token. lm_head is not used for that vector. It stays in the checkpoint so vLLM and SGLang can load the model without a serving error.

Scores

Hit@1. Parentheses are nDCG@5, and they appear only on Image, VisDoc, and Video. Image is the 36-task macro, visual documents the 27-task macro, and video the 18-task macro. SC and ARO are single scores. Both rows come from the same in-house evaluation.

	SC	ARO	Image	VisDoc	Video
WeMM-Embedding-2B, in-house	93.17	51.70	78.80 (87.34)	74.62 (78.08)	63.97 (78.74)
this model, in-house	92.92	49.00	76.99 (85.87)	73.49 (77.33)	62.28 (77.46)

Against WeMM-Embedding-2B, this model is lower by 0.25 points on SC, 2.70 on ARO, 1.81 on image Hit@1 (1.47 nDCG@5), 1.13 on VisDoc Hit@1 (0.75 nDCG@5), and 1.69 on video Hit@1 (1.28 nDCG@5).

Each ratio below is this model's score divided by the WeMM-Embedding-2B score on the same run. A ratio of 100% would match WeMM-Embedding-2B. Image, VisDoc, and Video use Hit@1. The nDCG@5 ratios are 98.3% for image, 99.0% for VisDoc, and 98.4% for video.

	SC	ARO	Image	VisDoc	Video
Kept vs WeMM-Embedding-2B	99.7%	94.8%	97.7%	98.5%	97.4%

MMEB-v2 tasks

Selected MMEB-v2 tasks, listed one by one rather than averaged again. Image and video scores are Hit@1. VisDoc scores are nDCG@5. All four columns come from in-house runs. WeMM-Embedding-2B and this model are from this run. Qwen3-VL-Embedding-2B and Eddy-VL 1.9B v1 are from an earlier in-house run of the same tasks.

Image (Hit@1)

Task	WeMM-Embedding-2B	this model	Qwen3-VL-Embedding-2B	Eddy-VL 1.9B v1
Visual7W-Pointing	98.60	98.60	82.40	84.00
RefCOCO	95.70	96.20	90.40	92.40
DocVQA	95.30	95.00	94.70	92.40
GQA	94.20	93.20	85.90	85.50
VOC2007	94.00	93.90	84.90	82.70
EDIS	93.30	92.10	88.30	82.00
RefCOCO-Matching	92.70	93.10	93.00	91.90
ImageNet-R	92.40	92.50	92.50	84.40
WebQA	91.70	91.00	88.60	85.60
TextVQA	91.50	90.80	88.30	85.00

Video (Hit@1)

Task	WeMM-Embedding-2B	this model	Qwen3-VL-Embedding-2B	Eddy-VL 1.9B v1
UCF101	99.30	97.20	80.10	65.00
Something-Something V2	91.30	89.30	68.10	59.50
HMDB51	88.80	84.50	71.90	60.30
ActivityNetQA	81.90	82.20	73.10	70.00
NExTQA	75.97	75.19	65.52	59.77
MSVD	72.54	71.49	68.96	66.87
QVHighlight	71.28	68.14	71.93	66.85

VisDoc (nDCG@5)

Task	WeMM-Embedding-2B	this model	Qwen3-VL-Embedding-2B	Eddy-VL 1.9B v1
ViDoRe · Synthetic DocQA (Healthcare)	98.89	98.52	95.28	93.93
VisRAG · SlideVQA	96.72	96.01	93.00	91.47
ViDoRe · Synthetic DocQA (AI)	96.32	95.95	96.72	94.62
ViDoRe · TabFQuAD	95.70	94.89	95.11	90.02
ViDoRe · Synthetic DocQA (Gov. reports)	95.47	93.48	96.45	93.81
ViDoRe · Synthetic DocQA (Energy)	93.91	93.89	90.28	89.94
VisRAG · InfoVQA	93.73	93.19	90.25	86.61
ViDoRe · InfoVQA	91.28	90.92	87.04	82.37
VisRAG · ChartQA	86.77	88.82	84.26	81.01
ViDoSeek · doc	83.60	83.48	84.79	82.21

Qwen3-VL-Embedding-2B and Eddy-VL 1.9B v1

Those two models come from a different family. Eddy-VL 1.9B v1 is distilled from Qwen3-VL-Embedding-2B. This model is distilled from WeMM-Embedding-2B, which keeps the Qwen3.5 hybrid backbone. Their scores in the tables above are from that earlier run.

The bars below count parameters used to produce an embedding. Eddy v1 is 1.93B and Qwen3-VL-Embedding-2B is 2.128B. In those two models, the token embedding and lm_head are one shared matrix, so the published size already counts that matrix once. WeMM is 2.213B and this model is 2.036B. Their lm_head is a separate matrix and is left out of the bar.

Size and compute

In the Gated DeltaNet stack, three linear-attention layers are removed. Softmax attention layers are not removed.

The embedding path does not use lm_head. Removing the three linear-attention layers reduces compute by 12% on long text and by 8.4% on a max-size image. Softmax attention and the vision tower are unchanged.

What is in this repo

File	Role
`model.safetensors`	BF16 weights, about 4.74 GiB, including `lm_head`
`config.json`	`WeMMEmbedding`, `trust_remote_code` via `auto_map`
`eddy_vl_embedding_v2.py`	Last non-pad token, then L2 normalization
`eddy_vl_embedding_v2_st.py`	Sentence-Transformers wrapper
`embedding_chat_template.jinja`	Appends the `<embedding>` token
`tokenizer.json`	Original WeMM tokenizer. It does not truncate at 8192. `model_max_length` is 262144

Not included: optimizer state.

Use

from huggingface_hub import hf_hub_download
from transformers import AutoModel, AutoProcessor

repo = "Urock-AI/eddy-vl_embedding_v2"
model = AutoModel.from_pretrained(repo, trust_remote_code=True, torch_dtype="auto").eval()
processor = AutoProcessor.from_pretrained(repo)
template = open(hf_hub_download(repo, "embedding_chat_template.jinja"), encoding="utf-8").read()
messages = [[
    {"role": "system", "content": [{"type": "text", "text": "Represent this input for retrieval."}]},
    {"role": "user", "content": [{"type": "text", "text": "a red apple on a table"}]},
]]
text = processor.apply_chat_template(
    messages, chat_template=template, add_generation_prompt=False, tokenize=False,
)
batch = processor(text=text, return_tensors="pt", padding=True)
print(model.embedding(**batch).shape)  # (1, 2048)

The processor is the original Qwen3VLProcessor. For a video, embedding_chat_template.jinja emits one <|video_pad|> from the frame list. Do not also put <|video_pad|> at the start of the question string.

For SGLang 0.5.9, follow the Serving section of WeMM-Embedding-2B. Run patch_sglang_video.py from that repository, then start the server with --is-embedding and --enable-precise-embedding-interpolation. That patch file is not in this repo. --model-path is this model, Urock-AI/eddy-vl_embedding_v2.

Citation

This model is a derivative of WeMM-Embedding-2B. The weights and code in this repository are released under Apache 2.0. Cite this model, and cite WeMM-Embedding as the base model.

@misc{eddy_vl_embedding_v2,
  title        = {Eddy-VL Embedding v2},
  author       = {Urock-AI Lab},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Urock-AI/eddy-vl_embedding_v2}}
}

@article{wemm-embedding,
  title         = {WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report},
  author        = {Junjie Zhou and Ke Mei and Lei Li and Tianyi Wang and Fengyun Rao and Jing Lyu},
  year          = {2026}
}