MMBU - Ungrounded Classification (Open-Ended): leaderboard
Open-ended classification on MMBU (2026), a biomedical image perception benchmark: the model names the finding in a whole image in free form, a Qwen3-32B judge checks strict biomedical equivalence, and the score is micro-averaged F1 times 100. Higher is better.
Metric: Open-ended (free-form) ungrounded classification of the whole image: micro-averaged F1 x100 (0-100) over MMBU's biomedical images from 35 submodalities (95% bootstrap intervals not published); zero-shot, official checkpoints; Qwen3-32B judges strict biomedical equivalence. Source: arxiv.org. Saturation forecast: Around October 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 13.7 | #585 |
| 2 | GPT-5.4 Mini | 10.4 | #389 |
| 3 | Qwen 2.5 VL 32B Instruct | 8.8 | #1053 |
| 4 | Qwen 3 VL 8B Instruct | 8.6 | #981 |
| 5 | Qwen 3 VL 4B Instruct | 7.9 | #1279 |
| 6 | GPT-4.1 Mini | 7 | #779 |
| 7 | MedGemma-4B-IT | 4.3 | #2716 |
| 8 | Qwen 2.5 VL 7B Instruct | 3.1 | #1833 |
| 9 | Gemma 3 4B (IT) | 2.2 | #3831 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mmbu-ungrounded-classification-open-ended · How It Works · Data refreshed daily, snapshot 2026-11-13.