MMTU - NL-2-code: leaderboard
The "NL-2-code" subtask of MMTU. Massive Multi-Task Table Understanding benchmark with real-world table reasoning tasks including joins, transformations, cleaning, and question answering.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around January 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 71.7 | #152 |
| 2 | GPT-5 | 71.1 | #98 |
| 3 | O3 | 71.01 | #124 |
| 4 | Gemini 2.5 Flash | 70.7 | #245 |
| 5 | GPT-5 Chat | 70.42 | #241 |
| 6 | GPT-5 Mini | 69.84 | #186 |
| 7 | Qwen 3 235B A22B 2507 (Thinking) | 69.45 | #260 (Qwen 3 235B A22B 2507) |
| 8 | Grok 3 Mini | 68.57 | #281 |
| 9 | DeepSeek R1 0528 | 68.11 | #222 |
| 10 | GPT-OSS-120B | 66.98 | #341 |
| 11 | Qwen 3 235B A22B 2507 Instruct | 65.39 | #300 |
| 12 | Llama 4 Maverick Instruct FP8 | 63.73 | #414 |
| 13 | GPT-OSS-20B | 62.34 | #518 |
| 14 | Qwen 3 32B | 61.5 | #443 |
| 15 | Qwen 3 8B | 60.43 | #674 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mmtu-nl-2-code · How It Works · Data refreshed daily, snapshot 2026-10-20.