Open Japanese LLM v2 - JFinQA: leaderboard
From LLM-jp's Open Japanese LLM Leaderboard v2, which runs open-weight models through llm-jp-eval v2 with vLLM. JFinQA poses numerical questions over Japanese corporate financial data, answered zero-shot as a boxed value and scored by mathematical equivalence with the reference. 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: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | llm-jp-4-32B-a3B-Thinking | 67.5 | #477 |
| 2 | GPT-OSS-20B | 66.5 | #523 |
| 3 | llm-jp-4.1-32B-a3B (Thinking) | 65.6 | #366 (llm-jp-4.1-32B-a3B) |
| 4 | llm-jp-4-8B-Thinking | 61.3 | #529 |
| 5 | GPT-OSS-120B | 60.9 | #346 |
| 6 | llm-jp-4.1-8B (Thinking) | 60.9 | #408 (llm-jp-4.1-8B) |
| 7 | llm-jp-4.1-33B (Thinking) | 60.2 | #256 (llm-jp-4.1-33B) |
| 8 | gemma-4-26B-A4B-it (Thinking) | 30.1 | |
| 9 | Qwen 3 8B (Thinking) | 27.9 | #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-jfinqa · How It Works · Data refreshed daily, snapshot 2026-11-05.