UnpredictaBench - Text: leaderboard

Metric: KS@100 (%): share of the 160 Text problems (named or described canonical distributions) for which 100 samples from the model (100 independent calls at temperature 1.0, reasoning effort none, up to 5 retries for a parseable answer) are not rejected by a two-sample Kolmogorov-Smirnov test against ground-truth samples at p < 0.0001; higher is better. Source: arxiv.org. Saturation forecast: Around 2039. 20 models tracked.

Top models

#ModelScore
1DeepSeek V3.2 (Non-reasoning)28.13
2GPT-5.4 (Non-reasoning)25.63
3Llama 3.1 70B Instruct23.13
4GPT-4o21.25
5Ministral-3-3B-Instruct-251221.25
6Nemotron 3 Nano 30B A3B (Non-reasoning)20.63
7Qwen 3.5 2B (Non-reasoning)14.38
8Qwen 3 32B (Non-reasoning)11.88
9Llama 3.2 1B Instruct10
10GPT-4o Mini9.38
11OLMo 3 7B Instruct7.5
12Grok 4.1 Fast (Non-reasoning)6.88
13Claude Sonnet 4.65.03
14Qwen 3.5 397B A17B (Non-reasoning)3.8
15Llama 3.1 8B Instruct3.75

Interactive version: theaggregate.ai/benchmark?slug=unpredictabench-text · How It Works · Data refreshed daily, snapshot 2026-09-29.