MKD Develops Proprietary Vision-Language Model Architecture KeuralVisionEncoder

MKD Co., Ltd. has unveiled KeuralVisionEncoder, its proprietary Vision-Language Model (VLM) architecture designed to enable AI systems to understand not only text, but also images and documents. The release represents another milestone in MKD's ongoing effort to build a fully independent multimodal AI ecosystem following the development of its bilingual language model, KeuralMoE-14B.

As multimodal AI rapidly emerges as the next frontier of artificial intelligence, the ability to process and reason across both visual and textual information has become increasingly important. While many existing vision-language models rely on publicly available vision backbones or third-party encoders, MKD has taken a different approach.

KeuralVisionEncoder was designed and developed entirely in-house, from the model architecture itself to the training pipeline. By building its own vision encoder rather than depending on external foundations, MKD aims to strengthen its technological independence and establish core competencies in multimodal AI research.

One of the key innovations incorporated into KeuralVisionEncoder is Adaptive Token Budgeting (ATB), a mechanism that dynamically allocates computational resources to the most important regions of an image. By focusing processing power where it matters most, the architecture is designed to improve both efficiency and visual understanding across a wide range of image types and complexities.

KeuralVisionEncoder is expected to become a core component of MKD's enterprise AI platform, Keural. Through this technology, users will be able to work not only with text, but also with images, charts, diagrams, scanned documents, and other visual information within a unified AI environment.

MKD is continuing research into next-generation vision models and larger-scale multimodal architectures. In the long term, the company plans to integrate KeuralVisionEncoder with its proprietary KeuralMoE language model family, creating a fully independent multimodal AI ecosystem optimized for enterprise applications.

The development of KeuralVisionEncoder demonstrates MKD's expansion beyond language models into the broader field of multimodal AI. The company remains committed to advancing its proprietary AI technologies and strengthening its position as a provider of sovereign AI solutions capable of competing in both domestic and global markets.