Nejumi 4 - JHumanEval: leaderboard
JHumanEval, a Japanese translation of HumanEval, run 0-shot in Nejumi Leaderboard 4 (Weights & Biases Japan). Score is the share of problems whose generated code passes all test cases in Nejumi's code-execution sandbox (code_exec_sandbox), 0-100; higher is better.
Metric: Sandbox pass rate (%). Source: nejumi.ai. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Max) | 94.87 | #16 (Claude Opus 5) |
| 2 | Phi-4 | 59.59 | #707 |
| 3 | GPT-4o (2024-11-20) | 58.77 | #392 |
| 4 | ABEJA-Qwen2.5-32B-Japanese-v1.0 | 55.13 | #516 |
| 5 | O3 Mini (2025-01-31) | 54.26 | #249 |
| 6 | GPT-4o Mini (2024-07-18) | 53.92 | #568 |
| 7 | Gemini 3.1 Flash Lite (Preview) | 53.44 | #201 |
| 8 | ELYZA-Shortcut-1.0-Qwen-32B | 52.64 | #560 |
| 9 | O4 Mini (2025-04-16) | 51.64 | #183 |
| 10 | Muse Spark Open (xHigh) | 50.13 | #42 (Muse Spark Open) |
| 11 | Gemma 4 31B (IT) | 50.1 | #178 |
| 12 | GPT-4.1 Mini | 49.74 | #363 |
| 13 | Gemini 3 Flash (Preview) | 48.82 | #90 |
| 14 | Hy3 (High) | 47.82 | #100 (Hy3) |
| 15 | Qwen 3.5 27B (Thinking) | 44.87 | #192 (Qwen 3.5 27B) |
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-jhumaneval · How It Works · Data refreshed daily, snapshot 2026-11-05.