NC AI unveils VAKEY Vision 3B model for robots and edge devices

NC AI said on November 3 that it has developed VAKEY Vision 3B, a 3-billion-parameter vision-language model aimed at physical AI, robots and industrial edge devices, according to aitimes.com.

The model processes video input in 8-frame units over 4 seconds, comparing relationships between frames to track robot motions and object trajectories. Its training data included industrial imagery such as welding defects, along with robot and spatial reasoning data.

In general response mode, VAKEY Vision 3B averaged 63.8 points across the BLINK, ERIQ and ERQA benchmarks, ahead of six open models including Qwen3-VL-2B/4B, Qwen3.5-2B/4B and Ministral-3-3B/8B. In reasoning mode it averaged 63.2 points across those three benchmarks plus RealWorldQA, against 48.4 points for Ministral-3-8B-Reasoning. NC AI said the model also scored highest among the compared models on ChartQA and TextVQA, which cover document understanding such as reading gauges, drawings and work instructions.

Welding-specific variant

NC AI fine-tuned the model on welding datasets to produce VAKEY Vision 3B Weld, which improved on the base SFT model on the RIAWELC, Khvatskii and RefWeld benchmarks.

The company said pre-training and fine-tuning were carried out on in-house infrastructure for security in domestic core industries and defense. Follow-up reinforcement learning is planned to embed the model as a vision-language-action model on robots and edge devices.

NC AI CEO Lee Yeon-su said VAKEY Vision 3B will be a key driver for deploying the company's robotics and VLA technology in the steel, shipbuilding and defense fields.