Naver Cloud said on Sunday it had unveiled HyperCLOVA X SEED 4B, a lightweight omnimodal model.

Naver Cloud showcased the model at the 2026 Korean Military Science and Technology Society comprehensive academic conference held in Jeju from June 10 to 12. It also unveiled a closed-network defence artificial intelligence roadmap, along with real-time analysis cases in edge environments such as drones and tactical vehicles.

HyperCLOVA X SEED 4B includes an in-house vision encoder, HyperCLOVA X CLIP, and an audio encoder to process images, video, voice and text simultaneously. It was trained on South Korea-focused data tailored to the country’s defence environment.

It optimised an existing 8B-class large language model backbone with pruning and knowledge distillation to halve the model size. Performance improved over the existing 8B omnimodal model, and low-latency inference is possible even in edge environments, Naver Cloud said.

Use cases include object detection based on drone and coastal surveillance video, change detection in satellite images, identifying risk factors in firing ranges and barracks, automatic recognition of military equipment, and battlefield map analysis.

Naver Cloud presented a defence AI full stack as its core vision that can be deployed and operated directly in a closed network, from infrastructure to MLOps, large language models and agents. It plans to build a defence AI foundation by this year and gradually complete AI transformation across all areas by 2030.