NC AI Unveils Four Multimodal Models in ‘VARCO-VISION 2.0’ Series

NC AI Leads the Domestic Multimodal AI Frontier with the Release of VARCO-VISION 2.0

South Korea’s leading AI company, NC AI, is making a bold leap forward in multimodal artificial intelligence with the release of VARCO-VISION 2.0, a suite of four multimodal AI models (14B / 1.7B / 1.7B OCR / Video-Embedding) designed for advanced image-text understanding. NC AI announced on the 16th that the models will be made available as open source, reinforcing its commitment to innovation and accessibility.

Built upon existing open-source text models and further trained with proprietary data, VARCO-VISION 2.0 showcases top-tier Korean language capabilities and the ability to jointly comprehend both images and text. This next-generation AI represents a major advancement in Korea’s sovereign AI technology.

The flagship VARCO-VISION 2.0 14B model surpasses state-of-the-art vision-language models such as InternVL3-14B, Alibaba's Ovis2-16B, and Qwen2.5-VL 7B in several key benchmarks, including English and Korean image understanding and OCR performance. Of the four models, the 14B and two embedding models are being released today, with the 1.7B base and OCR models scheduled for release next week.

What sets VARCO-VISION 2.0 apart is its ability to analyze multiple images simultaneously—making it especially effective at handling complex documents, tables, and charts. The models demonstrate high fluency in both Korean and English, and significant improvements have been made in text generation and cultural understanding, particularly for Korean content.

The 14B model, unveiled today, has outperformed previous leading models across multiple multimodal AI benchmarks, signaling the potential for sovereign AI leadership in the multimodal space.

To support both industrial and personal applications, NC AI is releasing a lightweight 1.7B parameter model alongside the 14B. While the 14B model is optimized for advanced reasoning and multi-image processing in enterprise settings, the 1.7B model is designed to run efficiently on personal devices such as smartphones and PCs—broadening the accessibility and scalability of multimodal AI.

The company also introduced VARCO-VISION-1.7B-OCR, a model tailored for optical character recognition (OCR) tasks. Unlike conventional OCR models, this model leverages a VLM-based approach that integrates image and language understanding, yielding significantly better performance in Korean OCR tasks.

One of the model's standout features is its AnyRes resolution-splitting input mechanism, which divides an image into segments and processes them at high resolution—enabling efficient handling of varied image sizes without loss of detail. This enables robust recognition even in noisy or blurry conditions, including mixed Korean-English environments.

The VARCO-VISION-Embedding model allows for the precise calculation of semantic similarity across text, image, and video data within a high-dimensional embedding space. This embedding technology enables video content to be indexed and searched using natural language queries, with related content retrieved based on vector similarity.

A notable innovation in this model is the integration of search vector transfer, which significantly enhances video search performance. By mathematically transferring the weight differences from pre-trained high-performance image-text models into fine-tuned video-text models, NC AI achieved state-of-the-art zero-shot performance on the MultiVENT2.0 video search benchmark—without additional fine-tuning.

These four new models are highly versatile, with applications across finance, education, culture, retail, and manufacturing. Use cases include automated document analysis, digitalization of contracts and invoices, structured data extraction from charts and tables, product image captioning, natural language video search, creative content generation, and advertising copywriting.

Beyond performance gains, VARCO-VISION 2.0 also sets new standards in computational and data efficiency. With refined data selection and novel data synthesis techniques, the models can be trained efficiently with reduced computational resources.

The VARCO-VISION-Embedding model, in particular, employs a novel technique to adapt existing preference-optimization datasets for contrastive learning, enhancing data efficiency and cost-effectiveness in model development. NC AI aims to make high-performance AI more accessible to developers and enterprises alike.

With this release, NC AI has reaffirmed its technical leadership by demonstrating capabilities not only in LLM development from scratch but also in advanced multimodal model construction. The models combine Korean language specialization with global-level performance, significantly advancing Korea's AI competitiveness.

All four models will be open-sourced for research purposes, available for use by individuals, enterprises, and public institutions alike. This move is seen as a major step in promoting a more inclusive and open AI ecosystem in Korea.

Through this release, NC AI seeks to enhance both the sovereignty and accessibility of Korea’s AI technology, contributing to the national agenda of strengthening sovereign AI. The practicality and high performance of these models are expected to serve as a catalyst for AI-driven innovation across multiple sectors in Korea.

Younsoo Lee, CEO of NC AI, stated,

“As AI technology evolves, the global trend is shifting beyond text-only language models toward vision-language models. With the release of these four models, NC AI reinforces its leadership in multimodal AI and affirms Korea’s potential to maintain sovereignty even in the era of vision-language AI.”

NC AI, a specialized AI subsidiary of NCSOFT, operates under the mission, “Everyone can be a Creator.” The company develops and provides AI solutions that drive creativity and business innovation across industries. Its AI technologies—including audio, graphics, translation, and chatbot solutions—are continually enhanced through in-house R&D and external collaborations, helping clients maximize productivity and creativity with tailored, industry-specific solutions.