Twinkle Eval - TW-Legal Benchmark v1: 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. TW-Legal Benchmark v1 has 209 four-choice questions on Taiwan law in Traditional Chinese, drawn from bar and certification exams; metric is accuracy (%), 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 | Llama 4 Maverick Instruct FP8 | 61.4 | #424 |
| 2 | Llama 3.3 70B Instruct | 50.38 | #541 |
| 3 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 49.36 | #219 |
| 4 | GPT-OSS-120B (Medium) | 49.12 | #346 (GPT-OSS-120B) |
| 5 | Mistral-Large-3-675B-Instruct-2512 | 48.87 | #274 |
| 6 | Llama 4 Scout Instruct | 47.21 | #547 |
| 7 | GPT-OSS-120B (Low) | 46.09 | #346 (GPT-OSS-120B) |
| 8 | GPT-OSS-120B (High) | 45.93 | #346 (GPT-OSS-120B) |
| 9 | Gemma 3 27B (IT) | 43.75 | #528 |
| 10 | NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 41.91 | #688 |
| 11 | Ministral-3-14B-Instruct-2512 | 38.29 | #625 |
| 12 | Gemma 3 12B (IT) | 35.2 | #672 |
| 13 | Llama 3.1 8B Instruct | 34.67 | #1017 |
| 14 | GPT-OSS-20B (High) | 32.7 | #523 (GPT-OSS-20B) |
| 15 | GPT-OSS-20B (Low) | 32.54 | #523 (GPT-OSS-20B) |
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-tw-legal-benchmark-v1 · How It Works · Data refreshed daily, snapshot 2026-11-05.