LiveMathBench 202412: leaderboard

Models answer 119 recent competition and entrance-exam math questions (CNMO, Chinese gaokao, AMC, Putnam), each in English and Chinese, 48 sampled answers per question; the score is mG-Pass@16, which rewards getting a question right consistently across draws rather than once. Qwen2.5-72B-Instruct checks each answer against the reference and is itself withheld. Unlike a pass@1 math set, a model that is right only on some draws scores low.

Metric: mG-Pass@16 on 238 questions (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 27 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1DeepSeek R177.6#487
2O3 Mini76.8#547
3DeepSeek R1 Distill Qwen 32B69.7#1806
4DeepSeek R1 Distill Llama 70B69.7#1669
5DeepSeek R1 Distill Qwen 14B68#2683
6O1 Mini64.8#776
7DeepSeek-R1-Distill-Qwen-7B63.1#7957
8QwQ 32B-Preview61.2#1628
9DeepSeek R1 Distill Llama 8B52.2#6937
10DeepSeek V347.9#734
11DeepSeek R1 Distill Qwen 1.5B45.1#9148
12Qwen2.5-Math-72B-Instruct42.3#1757
13Qwen 2.5 Max42.2#966
14Qwen 2.5 32B Instruct38.1#1257
15Qwen2.5-Math-7B-Instruct36.6#8400

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

Interactive version: theaggregate.ai/benchmark?slug=livemathbench-202412 · How It Works · Data refreshed daily, snapshot 2026-11-17.