GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
emirozcan16
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
emirozcan16
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Benchmark | AA-VL-4B |
|---|---|
| OSWorld-G Original | 65.3 |
| OSWorld-G Refined | 74.5 |
| UI-Vision Final | 29.93 |
UI-Vision breakdown: Basic 39.28 · Functional 37.75 · Spatial 14.21.
No cross-model comparison table here on purpose: published numbers for other models vary with prompt, protocol and agent scaffolding, so we report only our own raw runs and ship the artifacts to reproduce them.
Per-sample re-scoring artifacts ship with this repo — every raw generation, coordinate and ground-truth box is included:
| Benchmark | File | Rows | Correct |
|---|---|---|---|
| OSWorld-G (Original / Refined) | osworld_g_predictions.jsonl | 1128 (564 × 2) | 368 / 420 |
| UI-Vision Final | predictions.jsonl | 5479 | 1640 |
text
python3 -c "import json;r=[json.loads(l) for l in open('osworld_g_predictions.jsonl')];print({v:sum(a['correct'] for a in r if a['variant']==v) for v in ('original','refined')})"OSWorld-G rows store both the emitted 0–999 coordinates (pred_xy_norm999) and
the pixel coordinates used by the official rule (pred_xy_px); UI-Vision rows
store pred_xy_px + target_bbox_px. Every row carries the model output, the
ground-truth box and the computed correct, so both benchmarks can be re-scored
from scratch. See VERIFY.md.
python
from transformers import AutoProcessor, AutoModelForImageTextToTextfrom PIL import Image MODEL = "emirozcan16/AA-VL-4B"proc = AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)model = AutoModelForImageTextToText.from_pretrained( MODEL, trust_remote_code=True, torch_dtype="auto", device_map="auto") img = Image.open("screen.png").convert("RGB")msgs = [{"role": "user", "content": [ {"type": "image", "image": img}, {"type": "text", "text": "Instruction: Click the funnel icon right of the search bar. " "Reply with the single action line now, nothing else:"}]}]txt = proc.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)inp = proc(text=[txt], images=[img], return_tensors="pt").to(model.device)out = model.generate(**inp, max_new_tokens=64, do_sample=False)print(proc.batch_decode(out[:, inp["input_ids"].shape[1]:], skip_special_tokens=True)[0])# -> click x=<int> y=<int> (0-999 normalized scale) or refuseOptional LoRA-only weights: emirozcan16/AA-VL-4B-adapter.
For the exact prompt used in the reported scores, see SYS_CLICK in
osworld_g_eval.py; the snippet above is the minimal form.
~7K curated rows (~60% click, ~40% refusal). Sources:
| Source | What we used | Role | License |
|---|---|---|---|
| Jedi (xlangai/Jedi, HF) | icon/component grounding clicks (train_raw/val_raw, os_layout, component libraries) | click labels | see upstream |
| GroundCUA | GUI grounding click slice (~2.5K rows, picked_json + images) | click labels | MIT |
| ScreenSpot-v2 (OS-Copilot/ScreenSpot-v2, HF) | desktop screenshots only | build synthetic refusal pairs | see upstream |
| Our synthesis | instruction ↔️ unrelated screenshot cross-pairing (doc–doc, icon–icon, desktop) |
Base model: Qwen/Qwen3-VL-4B-Instruct (Apache-2.0). No private or scraped user data. The assembled training set is not released for now. We did not train on any benchmark's answer labels.
OSWorld-G_refined.json instructions, same 564 samples (510 click + 54 refusal), reported separately.Single greedy prediction per sample, no crop/zoom/planner/pass@k.
Built in ~3 months on a ~$8 compute budget. Contact: ozcannemir08@gmail.com
| refusal examples |
| ours |