FCMBench V1.1 - Numerical Calculation: leaderboard
Evaluates visual credit-document understanding on FCMBench V1.1: 26 document types, 5,198 images and 13,806 VQA pairs. Exact-match key/value sets are scored with F1; task scores macro-average fine-grained subtasks. Numeric calculation permits differences below 2. The paper prefers instruct or lowest available reasoning effort. The two-stage DeepSeek-OCR plus V3.2 system stays distinct from either model alone. These are paper V1.1 scores, not the differing repository leaderboard. This series reports FCMBench V1.1 - Numerical Calculation using F1 (%).
Metric: F1 (%). Source: arxiv.org. Saturation forecast: Around February 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 59.75 | #82 |
| 2 | Gemini 3 Flash | 57.26 | #103 |
| 3 | Kimi K2.5 | 55.87 | #152 |
| 4 | Qwen 3.5 35B A3B | 54.71 | #268 |
| 5 | Qwen 3.5 122B A10B | 54.64 | #184 |
| 6 | Qwen 3.5 397B A17B | 50.16 | #154 |
| 7 | Claude Opus 4.5 | 43.01 | #87 |
| 8 | Qwen 3 VL 235B A22B Instruct | 39.89 | #280 |
| 9 | GLM-4.5V | 39.7 | #358 |
| 10 | Qwen 3 VL 32B Instruct | 35.93 | #290 |
| 11 | GLM-4.6V | 32.32 | #330 |
| 12 | GPT-5.2 | 30.02 | #115 |
| 13 | Qwen 3 VL 8B Instruct | 29.98 | #419 |
| 14 | Llama 4 Maverick | 27.64 | #486 |
| 15 | Qwen 3 VL 30B A3B Instruct | 26.04 | #390 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=fcmbench-v1-1-numerical-calculation · How It Works · Data refreshed daily, snapshot 2026-11-05.