QIMMA - 3LM HumanEval+: leaderboard
TII's 3LM Arabic version of HumanEval+: Python function generation from Arabic instructions, checked against EvalPlus's extended test suites in the QIMMA Arabic leaderboard. pass@1 on a 0-100 scale, higher is better; open-weight models only.
Metric: pass@1 (%). Source: huggingface.co. Saturation forecast: Around February 2027. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 27B | 86.59 | #182 |
| 2 | Qwen 3.5 122B A10B | 85.98 | #184 |
| 3 | Qwen 3.5 27B | 84.76 | #192 |
| 4 | Qwen 3 235B A22B 2507 Instruct | 84.76 | #315 |
| 5 | Qwen 3.5 397B A17B | 84.15 | #154 |
| 6 | Qwen 3.6 35B A3B | 83.54 | #238 |
| 7 | Qwen 3.5 35B A3B | 81.1 | #268 |
| 8 | Qwen 3.8 27B | 79.88 | #123 |
| 9 | Qwen 3 30B A3B 2507 Instruct | 78.66 | #492 |
| 10 | gemma-4-E4B-it | 77.44 | #577 |
| 11 | Qwen 2.5 72B Instruct | 75.61 | #469 |
| 12 | Gemma 3 27B (IT) | 75 | #528 |
| 13 | GPT-OSS-20B | 74.39 | #523 |
| 14 | Qwen 2.5 32B Instruct | 73.78 | #509 |
| 15 | Qwen 3.5 9B | 73.17 | #391 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=qimma-3lm-humaneval-plus · How It Works · Data refreshed daily, snapshot 2026-11-05.