Open FinLLM Reasoning - XBRL-Math: leaderboard
Open FinLLM Reasoning subtask requiring arithmetic and financial calculations over XBRL-tagged structured filings.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 86.67 | #251 |
| 2 | DeepSeek R1 Distill Llama 70B | 86.67 | #594 |
| 3 | QwQ-32B | 84.44 | #421 |
| 4 | DeepSeek R1 Distill Qwen 32B | 84.44 | #655 |
| 5 | DeepSeek R1 Distill Qwen 14B | 84.44 | #844 |
| 6 | Qwen2.5-Math-72B-Instruct | 83.33 | #717 |
| 7 | DeepSeek R1 Distill Llama 8B | 81.11 | #1291 |
| 8 | DeepSeek V3 | 76.67 | #321 |
| 9 | GPT-4.5 | 74.44 | #274 |
| 10 | O1 | 74.44 | #183 |
| 11 | GPT-4o | 72.22 | #343 |
| 12 | Llama 3.3 70B Instruct | 70 | #534 |
| 13 | Qwen 2.5 72B Instruct | 67.78 | #454 |
| 14 | Qwen 2.5 32B Instruct | 65.56 | #498 |
| 15 | Llama 3.1 70B Instruct | 63.33 | #561 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=open-finllm-reasoning-xbrl-math · How It Works · Data refreshed daily, snapshot 2026-10-20.