Twinkle Eval - MMLU: leaderboard
From Twinkle AI's TW-Eval leaderboard (Twinkle Eval), which queries models through an OpenAI-compatible API and asks each question three times with shuffled options, reading the answer letter from a box. MMLU covers 57 English subjects; score is mean per-subject accuracy (%), higher is better. Unparseable answers count as wrong, so models that do not box an answer can fall below the 25% chance rate.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 90.37 | #219 |
| 2 | GPT-OSS-120B (High) | 88.18 | #346 (GPT-OSS-120B) |
| 3 | GPT-OSS-120B (Medium) | 88.02 | #346 (GPT-OSS-120B) |
| 4 | GPT-OSS-120B (Low) | 88 | #346 (GPT-OSS-120B) |
| 5 | Llama 4 Maverick Instruct FP8 | 87.26 | #424 |
| 6 | GPT-OSS-20B (High) | 85.16 | #523 (GPT-OSS-20B) |
| 7 | GPT-OSS-20B (Medium) | 84.99 | #523 (GPT-OSS-20B) |
| 8 | GPT-OSS-20B (Low) | 84.98 | #523 (GPT-OSS-20B) |
| 9 | Llama 4 Scout Instruct | 82.31 | #547 |
| 10 | Llama 3.3 70B Instruct | 80.89 | #541 |
| 11 | Mistral-Large-3-675B-Instruct-2512 | 80.72 | #274 |
| 12 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 76.06 | #688 |
| 13 | Gemma 3 27B (IT) | 71.6 | #528 |
| 14 | Ministral-3-14B-Instruct-2512 | 70.2 | #625 |
| 15 | Llama 3.1 8B Instruct | 55.55 | #1017 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=twinkle-eval-mmlu · How It Works · Data refreshed daily, snapshot 2026-11-05.