CFMME - Information Extraction: leaderboard

CFMME (2026) is a Chinese financial multimodal evaluation set for vision-language models, built from financial documents, charts, seals, tables and exam questions. This board is field-level F1 in percent for extracting key fields from financial documents; higher is better.

Metric: Field-level F1 (%) of key information extracted from financial documents on CFMME (6,052 Chinese financial multimodal instances over eight financial image types), zero-shot, mean of three runs, the non-reasoning model of each family; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 VL 235B A22B Instruct93.43#541
2GLM-4.5VNon-reasoning89.03#756
3Qwen 3 VL 8B Instruct88.86#1043
4InternVL3.5-241B-A28BNon-reasoning83.1#739
5ERNIE 4.5 VL 28B A3B80.71#2133
6InternVL3.5-8BNon-reasoning80.29#1618
7Kimi-VL-A3B-Instruct75.46#2272
8GPT-4o66.88#754
9Llama 3.2 11B Instruct47.52#5213

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=cfmme-information-extraction · How It Works · Data refreshed daily, snapshot 2026-11-22.