Twinkle Eval - TMMLU+: leaderboard
Metric: Accuracy (%). Source: huggingface.co. 29 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 78.79 |
| 2 | Llama 4 Maverick Instruct FP8 | 78.38 |
| 3 | GPT-OSS-120B (High) | 74.78 |
| 4 | GPT-OSS-120B (Medium) | 74.64 |
| 5 | GPT-OSS-120B (Low) | 74.61 |
| 6 | Mistral-Large-3-675B-Instruct-2512 | 73.31 |
| 7 | Llama 4 Scout Instruct | 67.87 |
| 8 | GPT-OSS-20B (Low) | 66.93 |
| 9 | GPT-OSS-20B (Medium) | 66.59 |
| 10 | GPT-OSS-20B (High) | 66.26 |
| 11 | Llama 3.3 70B Instruct | 64.68 |
| 12 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 59.48 |
| 13 | Gemma 3 27B (IT) | 57.35 |
| 14 | Ministral-3-14B-Instruct-2512 | 55.46 |
| 15 | Gemma 3 12B (IT) | 47.17 |
Interactive version: theaggregate.ai/benchmark?slug=twinkle-eval-tmmlu-plus · How It Works · Data refreshed daily, snapshot 2026-09-19.