VISTA-Bench: leaderboard
Metric: Accuracy (%) over all 1,500 questions (300 perception, 300 reasoning, 400 multimodal-knowledge and 500 unimodal-knowledge items, question-weighted) with every question rendered as an image of text (LaTeX pipeline, 800-pixel width, standard fonts) and read through the vision encoder; 1,500 manually checked questions drawn from MMBench, SEED-Bench, MMMU and MMLU (1,462 multiple-choice, 38 open); zero-shot with VLMEvalKit default decoding, final answers recovered by a GPT-based extractor; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 80.3 | #54 |
| 2 | Qwen 3.5 122B A10B | 75.5 | #170 |
| 3 | GLM-4.6V | 71.3 | #309 |
| 4 | GLM-4.1V-9B (Thinking) | 64.1 | #457 (GLM-4.1V-9B) |
| 5 | GPT-5.2 | 57.8 | #105 |
| 6 | Qwen 3 VL 8B Instruct | 52.3 | #401 |
| 7 | Qwen 2.5 VL 7B Instruct | 51.7 | #643 |
| 8 | Qwen 3 VL 30B A3B Instruct | 50.8 | #365 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=vista-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.