Open Japanese LLM v2 - StructEval: leaderboard
From LLM-jp's Open Japanese LLM Leaderboard v2, which runs open-weight models through llm-jp-eval v2 with vLLM. This task is the English non-rendering subset of TIGER-Lab's StructEval (JSON, CSV, XML, YAML and TOML generation), run zero-shot; score is 0.2 times parse success plus 0.8 times required-key validation. Runs evaluated before the task was added to llm-jp-eval lack it. Upstream 0-1 scores are shown as 0-100, higher is better.
Metric: StructEval Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 85.74 | #346 |
| 2 | GPT-OSS-20B | 85.66 | #523 |
| 3 | Qwen 3 8B (Thinking) | 80.78 | #678 (Qwen 3 8B) |
| 4 | llm-jp-4.1-33B (Thinking) | 80.27 | #256 (llm-jp-4.1-33B) |
| 5 | llm-jp-4.1-8B (Thinking) | 79.03 | #408 (llm-jp-4.1-8B) |
| 6 | llm-jp-4-32B-a3B-Thinking | 78.63 | #477 |
| 7 | llm-jp-4-8B-Thinking | 77.91 | #529 |
| 8 | llm-jp-4.1-32B-a3B (Thinking) | 77.68 | #366 (llm-jp-4.1-32B-a3B) |
| 9 | gemma-4-26B-A4B-it (Thinking) | 73.26 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=open-japanese-llm-v2-structeval · How It Works · Data refreshed daily, snapshot 2026-11-05.