Open Japanese LLM v2 - JCoLA Out-of-Domain: leaderboard
From LLM-jp's Open Japanese LLM Leaderboard v2, which runs open-weight models through llm-jp-eval v2 with vLLM. JCoLA asks for a binary acceptability judgment on a Japanese sentence; the out-of-domain split holds theoretically significant judgments from linguistics journal articles, scored by exact match. 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: Accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | llm-jp-4.1-8B (Thinking) | 82.87 | #408 (llm-jp-4.1-8B) |
| 2 | llm-jp-4-8B-Thinking | 81.35 | #529 |
| 3 | llm-jp-4.1-32B-a3B (Thinking) | 81.35 | #366 (llm-jp-4.1-32B-a3B) |
| 4 | llm-jp-4-32B-a3B-Thinking | 80.12 | #477 |
| 5 | llm-jp-4.1-33B (Thinking) | 79.51 | #256 (llm-jp-4.1-33B) |
| 6 | GPT-OSS-120B | 77.37 | #346 |
| 7 | GPT-OSS-20B | 70.95 | #523 |
| 8 | gemma-4-26B-A4B-it (Thinking) | 70.64 | |
| 9 | Qwen 3 8B (Thinking) | 68.5 | #678 (Qwen 3 8B) |
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-jcola-out-of-domain · How It Works · Data refreshed daily, snapshot 2026-11-05.