CorpusQA (128K): leaderboard
Corpus-level analysis benchmark (CorpusQA, Lu et al., Tongyi Lab, 2026) at its 128K-token scale: each question comes with about 92K tokens of concatenated real documents (financial reports in Chinese and English, education and real-estate documents), and its reference answer is computed by SQL over fields extracted from every document, so the evidence is spread across the corpus. A DeepSeek-V3 judge checks semantic equivalence with the reference; the score is the mean of the four domain accuracies in percent over 329 questions, and higher is better.
Metric: Accuracy (%; mean of the four domain accuracies, 128K-token corpora, DeepSeek-V3 judge). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 82.02 | #163 |
| 2 | Gemini 2.5 Pro | 80.19 | #262 |
| 3 | GPT-5 Mini (2025-08-07) | 79.88 | #341 |
| 4 | Qwen 3 235B A22B 2507 (Thinking) | 78.08 | #460 |
| 5 | Gemini 2.5 Flash | 77.81 | #452 |
| 6 | DeepSeek R1 0528 | 74.72 | #425 |
| 7 | MiniMax M1 80k | 71.02 | #926 |
| 8 | Qwen 3 30B A3B 2507 (Thinking) | 69.59 | #759 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=corpusqa-128k · How It Works · Data refreshed daily, snapshot 2026-11-12.