Encyclo-K: leaderboard

Each question shows 8-10 textbook knowledge statements, some rewritten to be false, and the model picks which of 4-8 option combinations names exactly the statements asked for (the true ones or, in other questions, the false ones); the score is accuracy over 5,038 questions (3,286 English, 1,752 Chinese) across 11 disciplines. Unlike MMLU-style items that test one fact, every answer depends on judging several statements jointly.

Metric: Accuracy on all 5,038 questions (%). Source: encyclo-k.github.io. Saturation forecast: Around February 2027. 37 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 235B A22B 2507 (Thinking)51.77#469
2Qwen 3 235B A22B 2507 Instruct50.4#618
3DeepSeek V3.1 TerminusThinking49.78#440
4Kimi K2 090544.66#564
5Qwen 3 235B A22BThinking43.25#631
6QwQ-32B40.97#908
7Qwen 3 30B A3B 2507 (Thinking)40.57#765
8Qwen 3 32BThinking37.28#965
9Qwen 3 30B A3B 2507 Instruct36.72#1091
10Qwen 3 30B A3BThinking31.56#1155
11DeepSeek V3 (0324)30.85#722
12Qwen 3 14BReasoning30.79#1264
13Qwen 3 4B 2507 Instruct28.6#2215
14Qwen 3 8BThinking26.78#1868
15Qwen 2.5 32B Instruct25.31#1257

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

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