Back to the newsroom

Press Release

Korea Deep Learning's KDL Frontier tops OCRBench v2, ahead of Gemini and GPT

Korea Deep LearningPublished

Korea Deep Learning's KDL Frontier tops OCRBench v2, ahead of Gemini and GPT
Korea Deep Learning's document-specialized VLM, KDL Frontier, placed first overall in the English category of OCRBench v2 in March 2026 with 68.1 points.

KDL Frontier, the document-specialized vision-language model (VLM) developed in-house by Korea Deep Learning, ranked first overall in the English category of OCRBench v2, a global multimodal optical character recognition (OCR) benchmark. In the March 2026 evaluation, KDL Frontier scored 68.1 overall, ahead of models from global AI companies such as Google Gemini, OpenAI and NVIDIA, and Korea Deep Learning announced the result on April 22, 2026. The company said it was the first time an Asian company had beaten Google Gemini to take the top spot in this evaluation.

A benchmark that evaluates document understanding across 31 scenarios

OCRBench v2 is a multimodal OCR benchmark listed on Hugging Face. Beyond the ability to read text, it evaluates a model's document understanding across 31 scenarios, including layout analysis, interpreting formulas and charts, and logical reasoning. It uses a dataset of 10,000 items verified by hand together with a separate private test set, and is designed so that models fitted only to public data find it hard to score well.

KDL Frontier led Gemini 3 Pro Preview, in second place, by 4.7 points, and also scored higher than major global models including GPT-5. The result shows that a document-specialized model, rather than a large general-purpose model, can be more accurate on document understanding tasks.

Near-zero hallucination, reading the values in a document as they are

Korea Deep Learning named extracting information that fits the context as a strength of KDL Frontier. The model finds the values it needs by taking into account the structure of a document, the relationships between fields and positional information together. It was also designed to extract the values written in a document as faithfully as possible, reducing hallucination. The company calls this "near-zero hallucination." The model was optimized to recognize text reliably even from noisy input such as Korean handwriting.

The result is a report card on Korea Deep Learning's strategy of training its own model on hundreds of millions of document records over five years. It is evidence of what document-specialized models can do in fields such as finance, legal and the public sector, where not a single wrong value is acceptable.

Where a first-place benchmark result reaches products and the field

KDL Frontier is the foundation model on which Korea Deep Learning's AI products stand. On April 2, 2026, Korea Deep Learning launched a SaaS version of DEEP Agent that can be used with a subscription alone, without building servers or infrastructure, and on April 7, 2026, it introduced DEEP Agent for Finance, specialized for financial document work. More than 80 customers, including public institutions and large companies, use Korea Deep Learning's AI products.

"In an AI market led by global big tech, it means something that an Asian company has surpassed Gemini and GPT with Korean technology alone. We will show the global market the value of hallucination-free document AI in industries such as finance, legal and the public sector, where high accuracy is required."

— Kim Dong-hyun, Chief Strategy Officer (CSO), Korea Deep Learning

About DEEP Agent

Korea Deep Learning makes AI that does not stop at reading documents. DEEP Agent is a document-based AI agent that interprets a document's visual structure and linguistic meaning together to extract the key information, then carries out the whole of the document work, from classification and verification to the hand-off into business systems. Upload a contract and it summarizes the key clauses, flags the risky ones and goes on to register it in the ERP, and it reaches the field in two weeks on average with no separate training. It holds GS Certification Grade 1, Korea's national software quality certification, and is designated for priority purchase by public institutions. More information is available on the DEEP Agent page at koreadeep.com.

About Korea Deep Learning

Korea Deep Learning (KDL), founded in 2019, is an enterprise AI agent company. Built on KDL Frontier, its own vision-language model (VLM) specialized for documents, KDL makes AI agents that take on and complete the document work of businesses and public institutions. Its main products are DEEP OCR, DEEP Parser, DEEP Agent and DEEP Ops.

Source