Jina VLM Release - WeMath: leaderboard
Jina-VLM technical report2512.04032v1, native jinaai/jina-vlm checkpoint. Author evaluation uses VLMEvalKit with English prompts matching training format; exact crop/output budgets and kit revision are not supplied. External comparator-paper scores are not pooled with native evaluations. WeMath. Native printed score retained on its declared scale. The second slash value marked dagger is the authors’ VLMEvalKit run of OpenGVLab/InternVL3_5-2B without thinking; the external first value is not imported.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3.5-2B (Non-reasoning) | 19.1 | |
| 2 | jina-vlm | 17.1 |
Interactive version: theaggregate.ai/benchmark?slug=jina-vlm-release-wemath · How It Works · Data refreshed daily, snapshot 2026-11-05.