VLAA-Thinker-Qwen2.5VL-3B: benchmark results
UC Santa Cruz VLAA's VLAA-Thinker Qwen2.5-VL 3B, a vision-language reasoning model built on Qwen2.5-VL-3B and trained with SFT/RL for R1-style reasoning, from their 'SFT or RL?' study. Provider: UCSC-VLAA. Released 2025-03-31. Access: Open.
Aggregate Rating 1519 ± 22, rank #726 of 1654 rated models, from 98 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open LMM Reasoning - MathVision - topology | 30.4 | Accuracy (%) | 89.3 |
| Open LMM Reasoning - DynaMath - Subject-solid geometry | 26.7 | Accuracy (%) | 82.7 |
| Open LMM Reasoning - WeMath - Angles and Length | 51 | Accuracy (%) | 82.1 |
| OpenVLM MathVista - STA | 80.1 | Accuracy (%) | 82 |
| OpenVLM MathVista - FQA | 70.6 | Accuracy (%) | 80.8 |
| Open LMM Reasoning - MathVerse - Analytic | 44.2 | Accuracy (%) | 79.6 |
| Open LMM Reasoning - MathVerse - Volume | 31.4 | Accuracy (%) | 79.6 |
| Open LMM Reasoning - WeMath - Correspondence of Coordinates and Positions | 80 | Accuracy (%) | 79.6 |
| Open LMM Reasoning - MathVision - analytic geometry | 36.9 | Accuracy (%) | 76.5 |
| OpenVLM MathVista - LOG | 27 | Accuracy (%) | 75.5 |
| Open LMM Reasoning - WeMath - InadequateGeneralization (Loose) | 13.5 | Accuracy (%) | 74.7 |
| Open LMM Reasoning - MathVerse - Applied | 39.4 | Accuracy (%) | 74.5 |
Interactive version: theaggregate.ai/model?slug=vlaa-thinker-qwen2-5vl-3b · How It Works · Data refreshed daily, snapshot 2026-10-24.