Docbench - Private Holdout: 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. 32 separately authored private holdout cases; published public-set column repeats main evaluation and is not duplicated. Difference is diagnostic; interval is uncertainty. Metric: Native task quality score (%). Source: https://bendyline.com/posts/docbench.html

Metric: Native task quality score (%). Source: bendyline.com. Saturation forecast: Around February 2027. 5 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.8 27B (Q4)82.6
2Gemma 4 12B81.3#931
3Qwen 3.5 4B71.2#1227
4MiniCPM5-2B65.6#1452
5Qwen 3.5 0.8B43.6#8136

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

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