Docbench - Structured Diagrams: leaderboard

Docbench in Gezel: fixed document-task adapters, parser/repair/retry behavior, native checks plus Claude Opus 5.5 rubric judge. Unusable outputs score zero. 32, 768 context except Apple 8, 192. Mostly Q4 is not affirmative precision evidence for each row; preserve explicit Q2/Q4 only. CLI/API labels are run configuration, not distinct public models. Native Apple Foundation Models row retained in pinned capture but not emitted until exact checkpoint identity is established. LLM-generated SVG/code and structured diagram outputs in the document workflow; source's native subject/composition rubrics and units, not image-model generation throughput. Metric: Native rubric score (0–100). Source: https://bendyline.com/posts/docbench.html

Metric: Native rubric score (0–100). Source: bendyline.com. Saturation forecast: Estimated already saturated. 10 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-OSS-20B91#1163
2Qwen 3.6 35B A3B91#484
3Gemma 4 12B91#931
4Qwen 3.8 27B (Q4)91
5Gemma 4 E4B89#2142
6Qwen 3.5 4B85#1227
7LFM3-1.7B85#2069
8Gemma 4 E2B76#4291
9Qwen 3.5 0.8B55#8136
10MiniCPM5-1B41#9079

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=docbench-structured-diagrams · How It Works · Data refreshed daily, snapshot 2026-11-17.