Tables#4
1. chandra-ocr-289.20
2. GLM-5.3-Flash89.18
3. Infinity-Parser2-Pro86.40
4. KDL-Frontier-Parser-nano85.56
5. dots.mocr85.20
Benchmarks · ParseBench by LlamaIndex
Ranked #1 of 33 in document parsing.
A public benchmark that comprehensively tests tables, charts, content fidelity, formatting and layout across 1,211 enterprise documents (about 2,000 pages) from the insurance, finance and public sectors. KDL's 1.2B parser first reached the top mean score in June 2026, and as of October 13 it still ranks #1.
76.36
Mean of 5 dimensions · Hugging Face leaderboard
Holding #1 · Oct 13, 2026 · first #1 in June 2026
#1Mean score among 33 models on the Hugging Face leaderboard
Top 5In tables, layout, formatting and charts at once. No other model does.
1.2BParameters. 23 times smaller than the 27B general model it outscores.
0Failures across 2,553 test cases in a single pass
Leaderboard · mean score
Ahead of models from the largest labs.
Ahead of models from Alibaba, Tencent, DeepSeek and NVIDIA, KDL's 1.2B parser took the top mean score.
1
KDL-Frontier-Parser-nano1.2B76.36
KDL-Frontier-Parser-nano1.2B76.362LightOnOCR-3-4BLightOn · 4.5B75.10
3LightOnOCR-3-0.8BLightOn · 0.9B74.60
4Infinity-Parser2-ProINF · 35.1B74.30
5Infinity-Parser2-FlashINF · 2.2B73.25
6MinerU2.5-Pro-2605-1.2BOpenDataLab · 1.2B72.78
7LightOnOCR-3-1BLightOn · 1.0B71.40
8
Qwen3.8-27B27.8B70.79
Qwen3.8-27B27.8B70.799GLM-5.3-FlashZhipu · 321B70.75
10chandra-ocr-2Datalab · 5.3B70.10
14
HunyuanOCR1.1B63.86
HunyuanOCR1.1B63.8616
DeepSeek-V4.1-Flash763B56.57
DeepSeek-V4.1-Flash763B56.5719
Nemotron-3-Nano-Omni-30B33.0B48.51
Nemotron-3-Nano-Omni-30B33.0B48.51Score vs. size
Small enough for
one server in a closed network.
Compared with giant models of hundreds of billions of parameters, our 1.2B small model achieved a higher parsing score. It can run as is on a single server in the closed networks of financial and public institutions.
By dimension
No single trick.
Strong across the whole page.
Models optimized for a single element see their accuracy drop sharply on real, mixed-document workloads. KDL's model ranks in the top five across all four areas: tables, charts, formatting and layout.
Layout · visual grounding#4
1. Infinity-Parser2-Flash80.61
2. TeleOCR80.14
3. MinerU2.5-Pro-2605-1.2B79.30
4. KDL-Frontier-Parser-nano78.84
5. PaddleOCR-VL-1.677.80
Semantic formatting#3
1. GLM-5.3-Flash79.49
2. LightOnOCR-3-4B67.60
3. KDL-Frontier-Parser-nano66.81
4. Qwen3-VL-8B-Instruct66.80
5. LightOnOCR-3-0.8B66.70
Charts#5
1. Qwen3.8-27B69.17
2. LightOnOCR-3-4B66.10
3. chandra-ocr-265.10
4. LightOnOCR-3-0.8B64.70
5. KDL-Frontier-Parser-nano63.41
Behind the model
How we built
a #1 parser at 1.2B.
In this interview, KDL's research team talks directly about how they designed and refined the 1.2B model to reach #1.
How a 1.2B Model Outparsed Google and OpenAI: Inside KDL-Frontier-Parser-nano
Read the interviewMethodology
How the score was measured
The benchmarkMeasured with LlamaIndex's ParseBench evaluation framework across 1,211 enterprise documents, about 2,000 pages in total, using 169,011 expert-reviewed rules.
Five dimensionsScored overall on table structural integrity, chart data values, missing or distorted text, bold and superscript formatting, and on-page position coordinates.
Single passUsing the official benchmark tool, ran 2,553 tests in a batch with a single parser model. There were zero inference failures.
Re-measuredLayout scores were re-measured by the ParseBench operators against the official evaluation metrics.
Open weightsThe weights of the default nano model are published on Hugging Face so the results can be verified directly. The high-performance ultra model is available through the KDL platform.
SourcesParseBench leaderboard · KDL-Frontier-Parser-nano · As of October 13, 2026
Benchmarks are a start.
Try it on your documents.
You can check parsing performance yourself with challenging document samples you use in practice. Your engineer will review the results with you.