DuckDB-NSQL: leaderboard
SQL generation benchmark using DuckDB: 75 questions across easy, medium, and hard difficulty testing natural-language-to-SQL translation with execution accuracy.
Metric: Execution Accuracy (%). Source: huggingface.co. Saturation forecast: Around January 2027. 99 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 Chat | 80 | #447 |
| 2 | Gemini 2.5 Flash (Preview 05-20) | 77.3 | #216 |
| 3 | Claude Sonnet 4.5 | 74.7 | #142 |
| 4 | GPT-4.1 | 74.7 | #248 |
| 5 | DeepSeek R1 | 74.7 | #251 |
| 6 | GPT-4o (2024-11-20) | 74.7 | #381 |
| 7 | Claude 3.7 Sonnet | 73.3 | #249 |
| 8 | O3 Mini | 72 | #272 |
| 9 | GPT-OSS-120B | 70.7 | #341 |
| 10 | Mistral Small 3 | 70.7 | #627 |
| 11 | O1 Preview | 70.7 | #193 |
| 12 | DeepSeek V3 (0324) | 69.3 | #340 |
| 13 | O1 | 69.3 | #183 |
| 14 | Claude 3.5 Sonnet | 68 | #350 |
| 15 | Qwen 2.5 Coder 32B Instruct | 68 | #528 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=duckdb-nsql · How It Works · Data refreshed daily, snapshot 2026-10-20.