Out of Set, a South Korea-based startup building AI models that run directly on hardware without internet connectivity, has closed a seed round led by The Ventures, a global early-stage venture capital firm. The investment will fund model training infrastructure, data acquisition, and key hires, according to Asia Economy Daily.
The company develops what it calls ultra-lightweight vertical AI models for on-device deployment. Its focus areas are speech recognition and speech synthesis, two capabilities in demand across security-sensitive industries including healthcare, legal, and finance, as well as smartphones, automobiles, and robotics.
Out of Set’s technical pitch targets a specific bottleneck in the current on-device AI pipeline. Most on-device models today start as large cloud-trained models that are then compressed through quantization and pruning before they can run on constrained hardware. That compression process can take months and often degrades performance. Out of Set trains models natively for edge device environments, producing models that can be integrated into products immediately without the performance loss that comes from compressing a cloud model down to fit.
The company has built an automated retraining system that adapts models to client requirements, allowing it to deliver customized solutions with fewer engineers and shorter timelines than traditional outsourcing approaches.
“Out of Set possesses the core technology to address the limitations of cloud AI at the device level,” Kim Chulwoo, CEO of The Ventures, said in the announcement. He added that the company is positioned to become “a key technology player” in the on-device AI market.
The Agent Infrastructure Angle
The investment arrives as the on-device agent infrastructure market accelerates. NXP Semiconductors shipped the first agentic AI framework for edge hardware at CES 2026 in January. Oppo open-sourced X-OmniClaw, a full Android agent framework that runs perception, memory, and action on the physical phone, on May 17. Perplexity, OpenClaw, and Hermes Agent have converged on Mac mini as reference hardware for persistent home-hosted agents.
Out of Set’s device-native training approach addresses the layer below these agent frameworks: the models themselves. If agents are going to run on phones, cars, and industrial edge devices, the models powering them need to be designed for those environments from the start, not squeezed down from cloud-scale architectures after training. That is the bet The Ventures is funding.