Glance-4B-preview
Calibrated typed decisions about an image or a text, in one forward pass.
GitHub (code, docs) · Usage · Benchmarks · Training data · License
Glance answers typed questions (find this object, pick an option, yes or no, rate on a scale) with calibrated probabilities. It reads the image or text once and answers every question from that single prefill, without generating tokens, and it can say "not present" with a probability attached. Eight questions about a COCO image take ~86 ms on an RTX PRO 6000, the same as one; the generative base model needs 3.2 s.
Real model outputs, rendered by examples/make_demo.py.
What is in this repository
| File | Content |
|---|---|
adapter_model.safetensors |
LoRA (r=64, alpha=128) on all linear layers of the Qwen3.5-4B language model, fp32 (122 M parameters) |
heads.safetensors |
box heads (presence, patch choice, 4 ordinal edge distributions, IoU) and the typed label readout, fp32 (12 M parameters) |
glance_config.json |
base model id, LoRA and head config, prompt limits, per-type temperatures |
The base weights are not included; they load from Qwen/Qwen3.5-4B
(Apache-2.0). The weights need the glance package: the shared-prefix forward and the heads are not part of
transformers.
Model description
Glance is a fine-tune of Qwen3.5-4B (LoRA r=64 plus small readout heads) that turns a vision-language model into a decision function. You give it a state (an image, a text, or both) and a list of typed questions; it returns a probability distribution per question:
| Primitive | You give | You get |
|---|---|---|
box |
a referring expression | P(present), a pixel box, P(IoU ≥ 0.5) and box confidence = P(present) · P(IoU ≥ 0.5) |
choice |
a question and 2–26 options | a probability per option |
yesno |
a yes/no question | P(yes) |
score |
a question and ordered levels | a probability per level and the expected level |
Why it is different from asking a chat model
- It can say "not present". The box primitive has a presence gate trained with explicit negatives (gRefCOCO no-target expressions and mined hard negatives). A generative grounder has to emit coordinates or a refusal string, and tends to draw a box anyway. On gRefCOCO no-target expressions Glance rejects 87.4% vs 65.0% for the base model.
- Calibrated. Probabilities come from proper-scoring-rule losses plus a per-type temperature fitted on held-out
training-distribution data. P(present) has 2.4% ECE and box confidence 3.8% ECE on gRefCOCO val, so a threshold
on
confidenceis a usable risk knob. - One pass, many questions. The state is prefilled once. Every question is a short block that continues from the prefix: the Gated-DeltaNet layers start from the prefix's recurrent state, and the full-attention layers see the prefix plus the block (bidirectional inside the block). Blocks never see each other, so answers do not depend on question order, and 8 questions cost about as much as 1.
- No parsing. The output is a distribution over the allowed answers, never free text.
How the box is decoded. A soft choice over the image's visual tokens (one per 32×32 px patch) picks the patch that holds the target; four ordinal distributions over 64 bins give the left, top, right and bottom edge offsets from that patch; the box is their expectation. An IoU head predicts whether that box will reach IoU ≥ 0.5.
Usage
pip install "glance @ git+https://github.com/Vision-Decision/Glance"
pip install flash-linear-attention # optional, CUDA only
from glance import Glance, Box, Choice, YesNo, Score
m = Glance.from_pretrained("Vision-Decision/Glance-4B-preview") # cuda, then mps, then cpu
m.box("park_bench.jpg", "the clear umbrella")
# BoxResult('the clear umbrella', present_prob=0.996, box=[189, 9, 374, 165], confidence=0.823)
m.box("park_bench.jpg", "a red bicycle") # not in the image
# BoxResult('a red bicycle', present_prob=0.000, box=[5, 171, 236, 363], confidence=0.000)
m.choice("park_bench.jpg", "Where are they?", ["park", "beach", "office", "train station"])
# ChoiceResult(answer='park', probs={'park': 0.994, 'beach': 0.004, 'office': 0.001, 'train station': 0.002})
m.yesno("editor.png", "Is a save dialog open?")
# YesNoResult(p_yes=0.992)
ticket = ("I was charged twice for my March invoice and the second payment is still pending. "
"Please refund the duplicate charge before Friday.")
m.score(ticket, "How urgent is this ticket?", levels=["low", "medium", "high", "critical"])
# ScoreResult(answer='high', expected=1.67, probs={'low': 0.048, 'medium': 0.371, 'high': 0.446, 'critical': 0.135})
# many typed questions in ONE shared-prefix pass; plain strings are box questions
rs = m.ask("vegetables.jpg", ["the broccoli", "a banana", YesNo("Is this food cooked?"),
Choice("What is the main food group?", ["vegetables", "meat", "dairy", "grains"])])
Outputs above are from an Apple-silicon Mac (MPS, fp32); the example images are in the
GitHub repository. Use max_vis_tokens=2048
for dense screenshots. Probabilities use the per-type temperatures fitted after training (presence 0.949,
choice 1.046, yes/no 0.889, score 1.046); pass calibrated=False for raw ones.
Benchmarks
Methodology and the script that renders this table: GitHub README.
Glance-4B-preview leads on absence and refusal, calibration and latency, and it is ahead of the generative base on referring grounding. It trails the GUI specialists clearly on ScreenSpot-Pro and VenusBench-GD grounding (and the base model on ScreenSpot-v2): the preview saw relatively little high-resolution GUI data, and it reads a screenshot in a single pass at ≤2,048 visual tokens without zoom.
Full results for Glance-4B-preview
| Benchmark | Metric | Glance-4B-preview |
|---|---|---|
| RefCOCO val / testA / testB | Acc@0.5 | 93.2 / 94.5 / 90.7 |
| RefCOCO+ val / testA / testB | Acc@0.5 | 89.7 / 93.3 / 84.6 |
| RefCOCOg val / test | Acc@0.5 | 90.0 / 90.4 |
| RefCOCO family, 8-split mean | Acc@0.5 | 90.8 |
| gRefCOCO val (no-target) | N-acc | 87.4 |
| gRefCOCO val + RefCOCO val | AUROC of P(present) | 97.9 |
| gRefCOCO val + RefCOCO val | ECE of P(present) / of box confidence | 2.4% / 3.8% |
| VenusBench-GD | refusal / grounding (positives) / overall | 81.5 / 47.3 / 52.3 |
| ScreenSpot-v2 | click accuracy | 89.4 |
| ScreenSpot-Pro | click accuracy | 29.3 |
| OpenDecisions (2,352 items) | capability / calibration | 67.6 / 80.4 |
| Latency, RTX PRO 6000, batch 1 | 1 / 8 box questions per image; 1 text question | 85.5 / 85.6 / 67.3 ms |
Comparison (same as the image)
| Category | Benchmark (metric) | Glance-4B-preview | Qwen3.5-4B | Qwen3-VL-8B | UI-Venus-1.5-8B | KV-Ground-4B | RC-GRPO-II-4B |
|---|---|---|---|---|---|---|---|
| Referring grounding | RefCOCO / + / g, average (Acc@0.5) | 90.8ᵃ | 88.1ᵇ | 89.1ᶜ | — | — | — |
| RefCOCO val (Acc@0.5) | 93.2 | 91.4ᵈ | — | — | — | — | |
| RefCOCO+ testB (Acc@0.5) | 84.6 | — | — | — | — | — | |
| RefCOCOg test (Acc@0.5) | 90.4 | — | — | — | — | — | |
| Absence / refusal | gRefCOCO val, no-target (N-acc) | 87.4 | 65.0ᵈ | — | — | — | 85.3ᵉ |
| VenusBench-GD refusal (acc.) | 81.5ᶠ | — | 63.8ᵍ | 61.6ʰ | — | — | |
| GUI grounding | ScreenSpot-v2 (click acc.) | 89.4 | 93.4ᵈ | 92.1ʰ | 95.9ʰ | 94.1ⁱ | — |
| ScreenSpot-Pro (click acc.) | 29.3 | 60.3ᵇ | 54.6ᶜ | 68.4ʰ | 67.0ⁱ | — | |
| VenusBench-GD overall (acc.) | 52.3ᶠ | — | 62.3ᵍ | 72.3ʰ | — | — | |
| Typed decisions | OpenDecisions (capability) | 67.6ʲ | 63.6ʲ | — | — | — | — |
| Calibration | Box-confidence ECE, gRefCOCO val (↓) | 3.8%ᵏ | — | — | — | — | — |
| P(present) ECE, gRefCOCO val (↓) | 2.4%ᵏ | — | — | — | — | — | |
| Efficiency | Latency, 1 question (ms, ↓) | 85ˡ | 400ᵈ | — | — | — | — |
| Latency, 8 questions per image (ms, ↓) | 86ˡ | 3,161ᵈ | — | — | — | — |
"—" means not reported in a primary source we could verify. Bold marks the best value in a row where at least two models report one.
Protocol
- Glance numbers come from one evaluation run of the released checkpoint (step 3,844) on an RTX PRO 6000
Blackwell in bf16 with the
flakernels. Every number is single-pass: no zoom, no test-time augmentation, no second look. Natural images are resized to ≤640 visual tokens (COCO 640×480 → 29×22 patches); GUI screenshots to ≤2,048 visual tokens (upscaling up to 2× allowed). - REC metric. Acc@0.5 is the fraction of expressions whose decoded box reaches IoU ≥ 0.5 with the ground truth, ignoring the presence gate (standard REC). With the gate required (P(present) ≥ 0.5), RefCOCO val drops from 93.2 to 92.0.
- GUI metric. Click accuracy: the center of the predicted box lies inside the ground-truth box.
- Calibration. 15 equal-width bins. Temperatures were fitted on held-out splits of the training distribution, never on an evaluation set.
- Decontamination. Every image that appears in any RefCOCO/+/g or gRefCOCO evaluation split was removed from training (RefCOCOg-umd train overlaps RefCOCO-unc val/test images), and evaluation images of the other benchmarks were removed by perceptual hash.
- External numbers are copied from the cited primary sources. Their prompts, resolutions and modes differ from ours (some report thinking mode, some use their own reproductions), so treat cross-column comparisons as indicative.
Footnotes
- ᵃ Mean of the 8 splits listed under Full results.
- ᵇ Qwen3.5-4B official model card, https://huggingface.co/Qwen/Qwen3.5-4B, rows "RefCOCO(avg)" and "ScreenSpot Pro". The card does not say which RefCOCO splits are averaged or whether thinking mode was used (thinking is the model's default).
- ᶜ Qwen3-VL technical report, arXiv:2511.21631, Table 4, Qwen3-VL-8B-Instruct ("RefCOCO-avg", splits not stated; ScreenSpot Pro).
- ᵈ Our measurement of the generative Qwen3.5-4B (bf16, greedy, thinking off, HF
generate) on the same GPU and image resolution as Glance. RefCOCO val and gRefCOCO no-target: random 500-expression subsets (seed 0) with the prompt "Locate the object described as … Output its bounding box as [x1, y1, x2, y2] … If no such object is in the image, answer exactly "none"". ScreenSpot-v2: all 1,272 items with the official ScreenSpot-Pro Qwen3.5 prompt at ≤4,000 visual tokens. Latency: median per image for 1 question, and for 8 questions about the same image answered by 8 sequentialgeneratecalls; unlike Glance's timing it includes the processor call (≈ a few ms) and it uses unoptimised HF generation, so a serving engine would be faster. - ᵉ RC-GRPO, arXiv:2608.04698, Table 1: RC-GRPO-II on Qwen3-VL-4B, gRefCOCO val N-acc. The same table gives 77.2 for Qwen3-VL-4B-Instruct.
- ᶠ Scored with our own script, not yet with the official UI-Venus evaluation code: a refusal item is correct iff P(present) < 0.5; a positive item is correct iff the box center lies inside the ground-truth box. Overall is the micro-average over all 6,165 items in our copy (899 refusal); the paper lists 6,166 items (900 refusal), and its Overall is the same micro-average including refusal items.
- ᵍ VenusBench-GD paper, arXiv:2512.16501, Table 5 (Qwen3-VL-8B: overall 62.34, refusal 63.78). The UI-Venus-1.5 report reproduces Qwen3-VL-8B much lower (overall 55.1, refusal 6.8), so refusal scores of generative models depend strongly on prompt and refusal format.
- ʰ UI-Venus-1.5 technical report, arXiv:2602.09082, Table 1 (ScreenSpot-Pro, ScreenSpot-v2, VenusBench-GD overall) and Table 9 (refusal). The Qwen3-VL-8B ScreenSpot-v2 value is that report's reproduction. The KV-Ground card lists UI-Venus-1.5-8B at 93.2 on ScreenSpot-v2.
- ⁱ KV-Ground-4B-BaseGuiOwl1.5-0228 model card, https://huggingface.co/vocaela/KV-Ground-4B-BaseGuiOwl1.5-0228 (GUI-Owl-1.5-4B base, no zoom).
- ʲ OpenDecisions' own
score.pyon all 2,352 items. Both models answer in one pass by reading the probabilities of the answer-label tokens (the base model through its LM head) with temperatures fitted on held-out training-distribution data. Without temperatures: Glance 67.1, base 51.9. - ᵏ gRefCOCO val (8,905 no-target expressions) merged with RefCOCO val (the same 1,500 images; this gRefCOCO release has no single-target val expressions), 25,063 expressions. Box confidence is scored against the event "present and IoU ≥ 0.5".
- ˡ Median over 100 RefCOCO val images, batch 1, bf16 +
fla, RTX PRO 6000 Blackwell Server Edition; includes the vision encoder and all heads, excludes JPEG decoding and resizing.
Training data
One epoch (3,844 steps, ~6.6 h on one RTX PRO 6000) over 212,373 states and 520,048 questions:
| Part | States | Questions | Content |
|---|---|---|---|
| Natural images | 7.6k COCO train2014 images | 114k box | RefCOCO / RefCOCO+ / RefCOCOg expressions; gRefCOCO no-target expressions; mined negatives (colour swaps, same-category swaps, cross-category swaps), 28% absent |
| GUI screenshots | 20k | 130k box | OS-Atlas desktop (Linux, macOS) and mobile (UIBert, widget captioning, RICO-SCA, aw_mobile) element grounding; AGUVIS GUIEnv web clicks (point targets); 10% absent |
| PixMo | 5.8k | 60k box/point, 8k score | PixMo-Points (point targets) and PixMo-Count (counts as score questions) |
| Text decisions | 179k | 158k choice, 28k yes/no, 22k score | tasksource procedural-typed-decisions, other typed-decision collections, and public NLP datasets recast as choice / yes-no / score (MMLU, HellaSwag, WinoGrande, GSM8K, WANLI, DynaSent, LogiQA and others) |
Rows marked non-commercial in the source metadata and rows derived from evaluation benchmarks were dropped from the text collections. Images of every RefCOCO/+/g and gRefCOCO evaluation split were removed from training, and images of the other evaluation benchmarks were removed by perceptual hash. The training data is not released yet.
Limitations
- Preview. One epoch on ~212k states / 520k questions. Expect rough edges outside the training mix.
- GUI grounding is not competitive yet, especially on high-resolution professional software (ScreenSpot-Pro 29.3). Use a GUI specialist if you need click accuracy; Glance's GUI strength today is state checks and refusal.
- One box per expression. The box head returns one box; for multi-target expressions it reports presence and the most likely instance.
- Calibration is measured in distribution. Temperatures were fitted on held-out training-distribution data. On shifted data (e.g. VenusBench-GD, AUROC of P(present) 0.74) probabilities are less reliable; recalibrate on your data if you depend on thresholds.
- Score questions are the weakest typed primitive (OpenDecisions score competence 42.6). Distributions over ambiguous rubrics are often flat, which is honest but less useful.
- English only was trained and evaluated.
- The box for an absent target is meaningless. Always gate on
present_proborconfidence. - Off CUDA it is not fast. Without
flash-linear-attentionthe GDN layers use the PyTorch reference kernel; a Mac takes ~1.4 s (MPS) to ~6 s (CPU) per image.
License
- These weights (adapter and heads): CC BY-NC 4.0, non-commercial use only. The training data includes gRefCOCO, which is licensed for non-commercial use, and some text sources whose licenses we could not confirm, so we release the weights under a non-commercial license.
- The base model Qwen/Qwen3.5-4B is Apache-2.0 and is downloaded from its own repository.
- The inference code (GitHub) is Apache-2.0.
Citation
A paper is in preparation. Until then:
@misc{glance2026,
title = {Glance: Calibrated Typed Decisions in One Forward Pass},
author = {Vision-Decision},
year = {2026},
url = {https://github.com/Vision-Decision/Glance},
note = {Glance-4B-preview, https://huggingface.co/Vision-Decision/Glance-4B-preview}
}


