SRD Reference - AIME 2025 (non-thinking): leaderboard
Table 2 Base rows only: the untouched Qwen3 checkpoint before any post-training method, native chat template with thinking disabled, zero-shot, 32,768-token generation budget, temperature 0.7, top-p 0.8. AIME 2025, avg@12 over 12 samples per problem. GRPO, TTRL, OPSD and SRD rows are separately trained policies and are not published. Metric: avg@12 (%). Higher is better. Source: https://arxiv.org/html/2610.22080v1
Metric: avg@12 (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 8B (Non-reasoning) | 21.39 | #671 (Qwen 3 8B) |
| 2 | Qwen 3 4B (Non-reasoning) | 20.83 | #828 (Qwen 3 4B) |
| 3 | Qwen 3 1.7B (Non-reasoning) | 8.61 | #1187 (Qwen 3 1.7B) |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=srd-reference-aime-2025-non-thinking · How It Works · Data refreshed daily, snapshot 2026-10-18.