KV-Ground-4B-BaseQw3vl

KV-Ground-4B-BaseQw3vl

A small GUI grounding model optimized for high-resolution images.

• Type: Vision-Language Model (VLM) for GUI grounding
• Size: 4B parameters
• Input: Image + natural language instruction
• Output: Text
• Fine-tuned from: Qwen3-VL-4B-Instruct
• License: CC BY-NC-SA 4.0
• Developed by: Kingsware & Vocaela AI

<p align="left"> <a href="https://github.com/vocaela/kv-ground"><img src="https://img.shields.io/badge/GitHub-Repository-green?logo=github" alt="GitHub"></a> <a href="https://huggingface.co/vocaela/KV-Ground-4B-BaseQw3vl"><img src="https://img.shields.io/badge/Hugging%20Face-Model-orange?logo=huggingface" alt="Hugging Face"></a> </p>

Model Description

This model is developed to optimize small VLM models for high-resolution GUI grounding. We synthesize high-quality high-resolution GUI grounding data, and continue post-training Qwen3-VL-4B-Instruct with SFT followed by RFT (GRPO). Without reasoning CoT, it achieves 63.2 on ScreenSpot-Pro, becoming one of the best model at 4B range. Meanwhile, it maintains excellent performance on regular-resolution tasks with 94.6 on ScreenSpot-V2.

Key receipe:

• Data clean by MLLM as judge: various public GUI grounding datasets are of ~30% errors which highly degenerate model performance on high-resolution images. And hence we carefuly do multiple-rounds of data cleaning by MLLM-as-judge.
• Synthesize high-resolution GUI grounding data
• Continue post-training model through SFT and GRPO

---

Benchmark Results

• Impact of continue post-training on base models

  For the purpose of controlled comparison, all these numbers are re-/produced by us, using the same evaluation code in the repo kv-ground. The baseline numbers may be different from the sources. Please see below note section for the controlled setup.

  Models	ScreenSpot-Pro	ScreenSpot-V2	OSWorld-G	OSWorld-G-refined	UI-Vision
  Base: Qwen3-VL-4B-Instruct*	59.5	93.1	63.3	71.1	30.4
  KV-Ground-4B-BaseQw3vl*	63.2 (+3.7)	94.6 (+1.5)	64.0 (0.7)	71.2 (+0.1)	32.6 (+2.2)

  The results tell us:

  • Our continuous post-training method brings in consistent improvements
  • The high-resolution optimized training doesn't harm regular resolution tasks

• Comparision with top models under 8B (ranked by ScreenSpot-Pro)

  We consider the top models under 8B from ScreenSpot-Pro leaderboard and related most recent technical reports.

  Models	ScreenSpot-Pro	ScreenSpot-V2	OSWorld-G	OSWorld-G-refined	UI-Vision
  Specialized GUI Models
  UI-Venus-1.5-8B	68.4	93.2	69.4	-	-
  KV-Ground-4B-BaseGuiOwl1.5*	66.5	94.3	62.8	69.1	32.2
  MAI-UI-8B	65.8	95.2	60.1	68.6	40.7
  GUI-Owl-1.5-4B-Instruct*	65.3	92.8	61.7	66.8	30.4
  KV-Ground-4B-BaseQw3vl*	63.2	94.6	64.0	71.2	32.6
  Step-GUI-8B	62.6	95.1	70.0	-	-
  Step-GUI-4B	60.0	93.6	66.9	-	-
  Holo2-8B	58.9	93.2	70.1	-	-
  Holo2-4B	57.2	93.2	69.4	-	-
  GUI-Owl-7B	54.9	92.8	55.9	-	-
  OpenCUA-7B	50.0	92.3	55.3	-	29.7
  UI-Venus-1.0-7B	50.8	94.1	54.6	61.7	36.8
  GTA1-7B	50.1	92.4	60.1	67.7	-
  UI-TARS-1.5-7B	35.7	91.6	52.8	64.2	-
  General VLMs
  Qwen3-VL-4B*	59.5	93.1	63.3	71.1	30.4
  Qwen3-VL-8B	54.6	-	58.2	-	-

Notes:

• By default numbers are copied from each source
• * indicates the results produced by us
• For all the runs produced by us, for fair comparison, the same prompt structure of system -> user-image -> user-instruct is used. Similarly, the same system message is used, which is the default Qwen3-VL computer-use format prompt and also adopted by the ScreenSpot-Pro leaderboard. For OSWorld-G and OSWorld-G-refined, minor modification is made to instruct the refusal setting.

Quickstart

The model keeps the exact same architecture and configs of Qwen3-VL-4B-Instruct and hence the usage is the same. For detail examples and the grounding prompt, please checkout the benchmark evaluation code in the kv-ground repo.