ZoomGain Release Evaluation - Zoom Bench: leaderboard
ZoomGain Release Evaluation - Zoom Bench: Native benchmark accuracy or held-out macro mean (%). ZoomGain arXiv 2611.14076v1: ZoomGainBench bilingual fine-grained perception scores, weighted within-language by category counts. Table 3 author experiments train Qwen3.5-4B and 9B on 9K ZoomGain samples; official inclusionAI release identifies them as ZoomGain-4B/9B. ZwZ-retrained controls do not identify released checkpoints and are held. Table 3 Gemini-3-Flash (51.27) conflicts with Table 2 Gemini-3.6-Flash; that row is held, never inferred equivalent. Qwen3-VL labels are held absent Instruct/Thinking binding. Other source-named released models retain their exact labels and unreported effort.
Metric: Native benchmark accuracy or held-out macro mean (%). Source: arxiv.org. Saturation forecast: Around November 2028. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.3 Flash | 56.8 | #119 |
| 2 | Qwen 3.5 397B A17B | 54.91 | #242 |
| 3 | Qwen 3.5 122B A10B | 51.83 | #329 |
| 4 | Qwen 3.5 9B | 50.3 | #803 |
| 5 | Qwen 3.5 4B | 48.88 | #1180 |
| 6 | GPT-5.1 | 47.22 | #225 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=zoomgain-release-evaluation-zoom-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.