Hebrew LLM - Winograd (0-shot): leaderboard
The Winograd task of the Hebrew LLM Leaderboard — zero-shot Hebrew Winograd-schema pronoun disambiguation testing commonsense coreference; accuracy (%).
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 96.04 | #66 |
| 2 | Gemini 3 Flash (Preview) | 94.6 | #82 |
| 3 | GLM-5.3 Flash | 93.88 | #65 |
| 4 | Qwen 3.5 122B A10B | 93.17 | #173 |
| 5 | GLM-5.1 | 91.37 | #97 |
| 6 | GPT-5.1 | 91.01 | #135 |
| 7 | Gemini 2.5 Pro | 90.65 | #152 |
| 8 | GPT-4o | 89.21 | #343 |
| 9 | Claude Sonnet 4.5 | 88.85 | #142 |
| 10 | GPT-5 Mini | 88.49 | #186 |
| 11 | Qwen 3.5 397B A17B | 86.69 | #146 |
| 12 | Gemini 2.5 Flash | 85.97 | #245 |
| 13 | Gemma 4 31B (IT) | 85.97 | #176 |
| 14 | c4ai-command-a-03-2025 | 84.89 | #451 |
| 15 | Llama 3.3 70B Instruct | 83.45 | #534 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hebrew-llm-winograd-0-shot · How It Works · Data refreshed daily, snapshot 2026-10-20.