NC AI Unveils Compact Vision Model for Physical AI Robots

Benchmark results

The evaluation used BLINK, ERIQ and ERQA. In standard response mode, which answers without extended reasoning, Vacky Vision 3B beat six open models in the 2–8 billion parameter range, including Alibaba Qwen and Mistral-family models. In a step-by-step reasoning mode it scored 63.2, which is 14.8 points above the 48.4 of the 8B-class Ministral 3 8B Reasoning. The company stressed that these figures reflect specific evaluation conditions and do not mean superior performance on every vision-language task.

The model also posted the highest scores among compared models on ChartQA and TextVQA, which NC AI plans to use for reading gauges, drawings and work orders. It processes eight frames from a four-second video to track movement and position changes.

Welding variant and industrial plans

NC AI also trained a welding-specific version, Vacky Vision 3B Weld. In internal tests, welding classification accuracy rose from 41.1 to 80.9, defect detection and condition reasoning from 39.3 to 69.8, and weld location and shape recognition from 5.3 to 72.5. These are benchmark results, not validation on real production lines.

The company says it built its own pretraining and two-stage fine-tuning pipeline on its own infrastructure. It is developing AI for autonomous welding robots with Hanwha Ocean, running humanoid robot trials with POSCO DX and building a world model with Hyundai Rotem.

Next, NC AI plans to add reinforcement learning, deploy the model as a perception module in robots and industrial equipment, and extend it into a vision-language-action (VLA) model that also decides actions. CEO Lee Yeon-soo said high-performance lightweight models are essential for physical AI to work in real time on limited hardware.