Eddy-VL is a multimodal embedding model light enough to run on edge devices. It keeps the retrieval quality of a 2B-class vision-language embedder in a lighter, faster package — built at Urock-AI Lab for real-world multimodal search over images, video, and documents.
The story behind Eddy
At Urock-AI, our work is focused on one thing: solving the real problems that forensic investigators face.
And those problems start with the environment. Investigative work is closed by nature — often offline, frequently air-gapped, running on networks deliberately isolated from the outside world. An investigator rarely has a clean dataset and a datacenter. They have a seized drive, thousands of images and scanned documents, and a question that sounds simple: "find me everything that looks like this."
Most strong multimodal embedders can answer that question well, but they're heavy: slow to run, expensive to serve, and impossible to use where the data can never leave the room. That's exactly the case in forensics.
That's why, at Urock-AI, we build dedicated forensic hardware — including offline, edge-deployable devices meant to run right next to the evidence. Eddy-VL is designed for exactly that setting: we wanted the retrieval quality of a 2B-class model in something light enough to live on an edge device, with nothing leaving the room.
Rather than train a new model from scratch, we started from one of the best open multimodal embedders available and carefully slimmed it down, teaching the smaller model to stay faithful to the original's sense of what's similar to what. The result is Eddy-VL 1.9B: lighter and faster, while landing within roughly 10% of the original on the standard benchmark.
Why it might be a good fit
Runs on the edge. Light and fast enough to deploy directly on edge devices, no datacenter required.
Faithful, not just smaller. Tuned to preserve the original model's retrieval behavior rather than chase raw compression.
One space for everything. Images, video, text, and document pages all map into a shared embedding space, so you can search across them with a single query.
Korean-aware. Curated with Korean image–text data, for retrieval that isn't limited to English-only models.
What it is (at a glance)
Type
Multimodal embedding model (image · video · text · document → vector)
Size
1.93B parameters · 3.85 GB checkpoint
Speed
~1.1× faster than the base model
Quality
within ~10% of the base on MMEB-V2 (in-house, identical settings)
Embedding
2048-d, with shorter dimensions available for cheaper search
Context
up to 8192 tokens; flexible image resolution
How well it does
Eddy-VL is validated on MMEB-V2 across image, video, and document retrieval. Selected per-task results:
Image (hit@1)
Task
Score
DocVQA
92.4%
RefCOCO
92.4%
RefCOCO-Matching
91.9%
WebQA
85.6%
GQA
85.5%
TextVQA
85.0%
ImageNet-R
84.4%
Visual7W-Pointing
84.0%
VOC2007
82.7%
EDIS
82.0%
Video (hit@1)
Task
Score
ActivityNetQA
70.0%
MSVD
66.9%
QVHighlight
66.9%
UCF101
65.0%
HMDB51
60.3%
NExTQA
59.8%
Something-Something V2
59.5%
VisDoc (nDCG@5)
Task
Score
ViDoRe · Synthetic DocQA (AI)
94.6%
ViDoRe · Synthetic DocQA (Healthcare)
93.9%
ViDoRe · Synthetic DocQA (Gov. reports)
93.8%
VisRAG · SlideVQA
91.5%
ViDoRe · TabFQuAD
90.0%
ViDoRe · Synthetic DocQA (Energy)
89.9%
VisRAG · InfoVQA
86.6%
ViDoSeek · doc
82.2%
ViDoRe · InfoVQA
82.4%
VisRAG · ChartQA
81.0%
MMEB-V2 overall (mean across all datasets, in-house eval):
Model
Overall
Qwen3-VL-Embedding-2B (public leaderboard)
73.0
Qwen3-VL-Embedding-2B (our environment)
68.9
Eddy-VL 1.9B
63.2
The public leaderboard and our in-house pipeline differ in setup, so we compare Eddy against the teacher re-run in the same environment (68.9 → 63.2). That is roughly a 8% relative gap while being smaller and ~1.1× faster.
Fine-grained understanding (retained from the base model)
Capability retained vs. base 2B model
Compression usually costs compositional ability — telling "the door left of the shirt" from "the shirt left of the door." We checked this on four benchmarks that probe attribute / relation / binding and reasoning-driven retrieval, comparing Eddy against the model it was distilled from (in-house eval):
Benchmark
What it probes
Eddy-VL 1.9B
Base 2B
SugarCrepe
attribute / object / relation discrimination
86.1
86.4
ARO
attribute & relation matching
59.5
60.4
MR²-Bench (nDCG@10)
reasoning-intensive retrieval
24.5
24.7
Winoground (group)
text–image binding (hardest)
6.8
8.5
On attribute/relation and reasoning retrieval, Eddy stays within ~1 point of the base model — so the compression preserves fine-grained understanding, not just coarse retrieval. Winoground's group score (a deliberately hard 2×2 binding test where most models land in the single digits) shows the largest relative drop.
Clone the repo first, then run from inside the repo folder:
python
1import torch
2from PIL import Image
3from vl_embedding_v1 import VLEmbedder
45instruction ="Represent this input for retrieval."67# load from local repo checkout (weights in ./model.safetensors)8embedder = VLEmbedder(".", torch_dtype=torch.bfloat16, default_instruction=instruction)910# text / image / video → 2048-d vectors (L2-normalized)11text_vec = embedder.process([{"text":"a photo of a cat"}])[0]12image_vec = embedder.process([{"image": Image.open("photo.jpg")}])[0]13video_vec = embedder.process([{"video":"clip.mp4"}])[0]1415# cosine similarity = dot product16score =(text_vec @ image_vec.T).item()
You can also pass the Hub repo id ("Urock-AI/Eddy-vl_embedding_1.9B_v1") to download weights automatically, but you still need the cloned Python files on PYTHONPATH.
trust_remote_code=True is used internally for processing_vl.py (VLProcessor).
Good to know before you rely on it
It finds, it doesn't decide. Eddy-VL surfaces candidates for a human to review; it shouldn't be the sole basis for any high-stakes decision.
It's an embedder, not a reranker. For the final ordering of results, pair it with a dedicated reranker.
English benchmarks, broader goals. MMEB is English; multilingual and domain-specific performance aren't separately measured in this release yet (that's where we're headed next).
Smaller has a cost. A modest, expected quality dip versus the full base model comes with the smaller size.
What's next
This is the first step. Where we're taking Eddy-VL from here:
Broader language coverage — moving beyond English-centric evaluation toward strong retrieval across many languages.
Domain specialization — tuning for the kinds of fine-grained, reasoning-heavy search real-world workflows demand.
On-device efficiency — pushing further on size and speed so the model runs comfortably on edge hardware.
More to come.
Training data
Curated by Urock-AI Lab from public sources — including MS COCO, ko-coco (kms7530/ko-coco-bal), SUN, RVL-CDIP, CORD v2, and AI Hub multimodal datasets — alongside an internally curated image set.
Citation & license
Eddy-VL is derived from Qwen3-VL-Embedding-2B; please honor its license terms (Apache 2.0) and cite Urock-AI when you use this model. Evaluation uses the MMEB benchmark (Jiang et al., VLM2Vec, ICLR 2025).
