Benchmark
ScreenSpot Pro - Office
Vision benchmarks measure how well models understand images. Scores are percentages.
- Models measured
- 8
- Scores recorded
- 10
- Sources
- 1
- IQ weight
- 0.28
Score distribution
- 40.7%
- Mean
- 45.2%
- Median
- 31.4%
- Std dev
- 0.87%
- 10th percentile
- 81.7%
- 90th percentile
- 0.87%
- Min
This benchmark has a weight of 0.28 for chat models, measured over 8 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 1 | Holo2 30B A3BAgentic· ScreenSpot data | 83% |
| 2 | Holo2 30B A3B· ScreenSpot data | 81.7% |
| 3 | Qwen2.5 VL 72B Instruct· ScreenSpot data | 72.6% |
| 4 | Qwen 2.5 VL 32B Instruct· ScreenSpot data | 64.4% |
| 5 | Qwen2.5-VL 7B Instruct· ScreenSpot data | 45.2% |
| 6 | qwen2.5-vl-3b-instruct· ScreenSpot data | 29.6% |
| 7 | GPT 5resized, minimal· ScreenSpot data | 23% |
| 8 | GPT 5resized, high· ScreenSpot data | 6.1% |
| 9 | GPT 4o· ScreenSpot data | 0.87% |
| 10 | Qwen2-VL (72B) Instruct· ScreenSpot data | 0.87% |
Data credits: ScreenSpot grounding leaderboard (GUI-Agent), Apache-2.0 (licence)