Twinkle Eval - TMMLU+: 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. TMMLU+ is Traditional Chinese multiple choice across Taiwanese school and professional subjects; score is mean per-subject accuracy (%) over the subjects every run scored, higher is better. Unparseable answers count as wrong.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 78.79 | #219 |
| 2 | Llama 4 Maverick Instruct FP8 | 78.38 | #424 |
| 3 | GPT-OSS-120B (High) | 74.78 | #346 (GPT-OSS-120B) |
| 4 | GPT-OSS-120B (Medium) | 74.64 | #346 (GPT-OSS-120B) |
| 5 | GPT-OSS-120B (Low) | 74.61 | #346 (GPT-OSS-120B) |
| 6 | Mistral-Large-3-675B-Instruct-2512 | 73.31 | #274 |
| 7 | Llama 4 Scout Instruct | 67.87 | #547 |
| 8 | GPT-OSS-20B (Low) | 66.93 | #523 (GPT-OSS-20B) |
| 9 | GPT-OSS-20B (Medium) | 66.59 | #523 (GPT-OSS-20B) |
| 10 | GPT-OSS-20B (High) | 66.26 | #523 (GPT-OSS-20B) |
| 11 | Llama 3.3 70B Instruct | 64.68 | #541 |
| 12 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 59.48 | #688 |
| 13 | Gemma 3 27B (IT) | 57.35 | #528 |
| 14 | Ministral-3-14B-Instruct-2512 | 55.46 | #625 |
| 15 | Gemma 3 12B (IT) | 47.17 | #672 |
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-tmmlu-plus · How It Works · Data refreshed daily, snapshot 2026-11-05.