# BizFinBench leaderboard

Chinese business-finance queries drawn from real use, in five areas: numerical calculation, reasoning, information extraction, prediction recognition and knowledge QA (nine categories).

## BizFinBench: average score across 9 tasks (higher is better, 10 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | ChatGPT-o3 | OpenAI | 73.9 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 2 | DeepSeek-R1 (671B) | DeepSeek | 73.0 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 3 | GPT-4o | OpenAI | 71.8 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 4 | DeepSeek-V3 (671B) | DeepSeek | 71.6 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 5 | ChatGPT-o4-mini | OpenAI | 71.3 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 6 | Gemini-2.0-Flash | Google | 69.8 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 7 | Qwen2.5-VL-32B | Qwen | 68.6 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 8 | Qwen3-32B | Qwen | 68.3 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 9 | Qwen2.5-72B-Instruct | Qwen | 67.7 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |
| 10 | Qwen2.5-VL-72B | Qwen | 67.5 | average score across 9 tasks |  | Reported by the authors | Paper Table 3 | https://arxiv.org/html/2505.19457v1 | 26 May 2025 (checked 1 Nov 2026) |

Scores are never combined across benchmarks. Each row names its metric, source and date.

## About this benchmark

- Maker: HiThink Research
- Year: 2025
- Use case: Finance knowledge
- Size: 6,781 questions, 9 categories
- Public split: All 6,781 questions on Hugging Face (test split per subset)
- Grading: Accuracy for six tasks, with rule-based answer extraction. GPT-4o judge scores the other three (tool use, knowledge QA, data description). The paper proposes the IteraJudge framework.
- Licence: CC-BY-NC-4.0 (data, research use only); Apache-2.0 (code)
- Submission: maker-only
- Products allowed: No
- Latest version: v1 (paper 2025-05-26), marked outdated by the maker; BizFinBench.v2 released 2026-01-09 (29,578 questions, ICML 2026)
- What it does not cover: Chinese only in v1. No live retrieval, no agentic tool loops beyond one tool-usage task, and no English filings.
- Last checked: 1 Nov 2026
- Links: paper https://arxiv.org/abs/2505.19457, site https://hithink-research.github.io/BizFinBench/, data https://huggingface.co/datasets/HiThink-Research/BizFinBench, code https://github.com/HiThink-Research/BizFinBench
- Directory page: https://benchmarks.finance/benchmarks/bizfinbench/
- Leaderboard: https://benchmarks.finance/leaderboards/bizfinbench/

Source: https://benchmarks.finance/leaderboards/bizfinbench/

Created and maintained by Moneda.