VinCLAT (Catalan): leaderboard
Vinclat (Pàmies et al., 2026) is a Catalan riddle game of 1,000 human-written items: from four independent clues the model infers intermediate keywords, then connects them into a final answer that draws on Catalan culture and abductive reasoning. A riddle counts as solved only when the final answer matches exactly, and the score is accuracy (%) averaged over four prompts (English or Catalan instructions, letter-count or fill-in-the-blank format); higher is better. Qwen3 models were run with thinking on and with thinking off, and each run is a separate entry.
Metric: Average Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 28.25 | #251 |
| 2 | Qwen 3 235B A22B (Thinking) | 17.48 | #313 (Qwen 3 235B A22B) |
| 3 | DeepSeek V3 (0324) | 9.6 | #340 |
| 4 | Qwen 3 235B A22B (Non-reasoning) | 9.1 | #313 (Qwen 3 235B A22B) |
| 5 | Qwen 3 32B (Thinking) | 6.1 | #443 (Qwen 3 32B) |
| 6 | Qwen 3 14B (Reasoning) | 3.03 | #537 (Qwen 3 14B) |
| 7 | Qwen 3 30B A3B (Thinking) | 2.38 | #493 (Qwen 3 30B A3B) |
| 8 | Qwen 3 32B (Non-reasoning) | 1.78 | #443 (Qwen 3 32B) |
| 9 | Qwen 3 8B (Thinking) | 1.63 | #674 (Qwen 3 8B) |
| 10 | Qwen 3 14B (Non-reasoning) | 1.35 | #537 (Qwen 3 14B) |
| 11 | Qwen 3 30B A3B (Non-reasoning) | 1.3 | #493 (Qwen 3 30B A3B) |
| 12 | Qwen 3 4B (Reasoning) | 0.75 | #829 (Qwen 3 4B) |
| 13 | Qwen 3 8B (Non-reasoning) | 0.7 | #674 (Qwen 3 8B) |
| 14 | Qwen 3 4B (Non-reasoning) | 0.28 | #829 (Qwen 3 4B) |
| 15 | Qwen 3 1.7B (Thinking) | 0.15 | #1196 (Qwen 3 1.7B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vinclat-catalan · How It Works · Data refreshed daily, snapshot 2026-10-20.