Glance-X-4B finds objects in an image from a text request, such as "the kid with glasses in the middle" or "左边举起球拍的男孩". It draws a box around every matching object, or says the object is not in the image.
One set of weights runs in three directions, and you pick one for each call:
| Direction | Computed once and shared | Best for |
|---|---|---|
image_first image → text |
the image | one image, many requests, answered in parallel |
text_first text → image |
the request | one request over many images or video frames |
joint image ⇄ text |
nothing: image and request read each other | the most accurate answer for one pair |
Real outputs of this checkpoint on held-out images (MS COCO).
Use it
The code lives on GitHub: Vision-Decision/Glance-X.
pip install "glancex[fla] @ git+https://github.com/Vision-Decision/Glance-X" # drop [fla] without a CUDA GPU
from glancex import GlanceX
m = GlanceX.from_pretrained("Vision-Decision/Glance-X-4B") # this repo + Qwen/Qwen3.5-4B-Base
m.ask("tennis.jpg", ["the kid with glasses in the middle", "左边举起球拍的男孩", "a frisbee"]) # image_first
# Result('the kid with glasses in the middle', present_prob=1.000, box=[249, 165, 381, 424], confidence=0.997, n_boxes=1)
# Result('左边举起球拍的男孩', present_prob=0.992, box=[65, 143, 253, 422], confidence=0.987, n_boxes=1)
# Result('a frisbee', present_prob=0.000, box=[387, 93, 587, 330], confidence=0.000, n_boxes=0)
m.find("giraffes.jpg", "the three giraffes in the field", mode="joint").boxes
# [[47, 62, 242, 316], [333, 69, 515, 312], [522, 101, 638, 314]]
m.search("a giraffe eating from someone's hand", photos)[0] # text_first, ranked
# Result("a giraffe eating from someone's hand", present_prob=0.994, box=[307, 22, 585, 324], confidence=0.753, n_boxes=0)
Every result has these fields:
present_prob: P(the request matches something in the image)box: in pixelsiou_prob: P(IoU ≥ 0.5)confidence:present_prob × iou_probboxes: every target found by the multi-box head
Files
| File | Contents |
|---|---|
adapter_model.safetensors |
LoRA r=64 on every language-model linear layer of Qwen3.5-4B-Base (fp32, 488 MB) |
heads.safetensors |
presence, box (patch centre + 4×64-bin edges), IoU and multi-box objectness heads (59 MB) |
glancex_config.json |
model config (K=8, positions, head sizes) |
The base model Qwen/Qwen3.5-4B-Base is downloaded separately (about 9 GB). Its vision tower is used frozen.
How it works

- Modality stage (layers 1–8). The image and the request run as separate streams, so each can be cached.
- Fusion stage (layers 9–32). One stream becomes a shared prefix and the other a block that reads it. In
joint, both sit in one block with attention running both ways. - Heads. Presence, single box, IoU, and per-patch objectness for multi-target requests.
With K=0 and a causal block mask, every direction equals a plain causal forward of the equivalent single sequence; the exactness test checks this. More in ARCHITECTURE.md.
Results
All results are on the full splits. Each direction was evaluated separately with these weights.
Original benchmarks (human labels). RefCOCO family, Acc@0.5 (%):
| Glance-4B-preview (image-first only) |
image_first |
text_first |
joint |
|
|---|---|---|---|---|
| RefCOCO val / testA / testB | 93.2 / 94.5 / 90.7 | 92.9 / 94.8 / 90.4 | 92.1 / 93.9 / 89.5 | 93.1 / 94.9 / 90.5 |
| RefCOCO+ val / testA / testB | 89.7 / 93.3 / 84.6 | 89.2 / 93.0 / 84.7 | 87.6 / 91.6 / 82.0 | 89.6 / 93.1 / 84.9 |
| RefCOCOg val / test | 90.0 / 90.4 | 90.0 / 90.5 | 89.3 / 89.0 | 90.3 / 90.5 |
| 8-split mean | 90.8 | 90.7 | 89.4 | 90.9 |
gRefCOCO val, %:
| Metric | image_first |
text_first |
joint |
|---|---|---|---|
| AUROC of P(present) | 98.1 | 97.3 | 98.2 |
| multi-target F1 | 77.3 | 78.1 | 82.7 |
Glance-Data eval set (300 held-out images, 23,586 pipeline-labelled requests; in-distribution for this model):
| Metric | image_first |
text_first |
joint |
|---|---|---|---|
| Acc@0.5, English / Chinese | 88.1 / 86.8 | 87.2 / 85.9 | 87.8 / 86.8 |
| look-alike negatives rejected (attribute / relation) | 94.9 / 94.2 | 93.8 / 94.2 | 95.0 / 94.4 |
| multi-target F1 | 62.6 | 66.7 | 69.7 |
Full tables, including a stronger internal single-direction model that this checkpoint still trails by 1.5–3.8 points on RefCOCO, are in RESULTS.md.
Training
- Data. 355k original requests (RefCOCO / RefCOCO+ / RefCOCOg, plus gRefCOCO negatives) and 1.83M requests from
Glance-Data
(🤗 dataset). Glance-Data adds:
- short, long, casual and Chinese requests
- category-only and multi-target requests
- look-alike and count, position and near-category negatives
- Directions. Each batch uses one direction:
image_first50%,text_first25%,joint25%. - Schedule. 2 epochs = 6,100 steps at 8 images per step, LR 2e-4 (LoRA) and 5e-4 (heads), cosine schedule, bf16 with fp32 LoRA. This took about 3 h on one H200.
The training recipe is in TRAINING.md.
Limitations and license
- The model was trained on natural photos (COCO). It was not trained on screenshots or documents.
- The weights are licensed CC BY-NC-SA 4.0, for non-commercial use only, because the training data (gRefCOCO, Glance-Data) is non-commercial. The code on GitHub is Apache-2.0. The base model has its own license.
Citation
@software{glancex2026,
title = {Glance-X: one model, three ways to look},
author = {Vision-Decision},
year = {2026},
url = {https://github.com/Vision-Decision/Glance-X}
}
Model tree for Vision-Decision/Glance-X-4B
Base model
Qwen/Qwen3.5-4B-Base