NC AI's VAETKI Vision 3B is a tiny vision-language model built to run inside robots and factory gear

The model targets hardware with limited compute, such as robots and edge devices on factory floors. NC AI describes it as the vision engine for its future vision-language-action (VLA) models, which let robots understand their surroundings and the state of objects and then act on them[1][2].

For video, it analyses frames in 4-second, 8-frame windows at once, which NC AI says covers a wider time span than similar-sized models and helps the model pick up subtle robot movements and object trajectories. Its training data includes real industrial imagery, such as welding defects, along with robot and spatial reasoning data[1].

In its default mode, which skips step-by-step reasoning so it can answer quickly enough for real-time control, VAETKI Vision 3B averaged 63.8 across three robot and spatial understanding benchmarks: BLINK (multi-image spatial perception), ERIQ (embodied reasoning) and ERQA (embodied spatial reasoning). NC AI says that beats all six comparison models, Qwen3-VL-2B and 4B, Qwen3.5-2B and 4B, and Ministral-3-3B and 8B[1]. Its chart puts VAETKI at 70.1 on BLINK, 76.4 on ERIQ and 45.0 on ERQA.

In "Think mode", with deeper reasoning switched on, it scored 63.2, well ahead of Ministral-3-8B-Reasoning at 48.4. It also posted the top scores among the compared models on ChartQA (85.1) and TextVQA (84.3), which NC AI pitches as reading gauges, drawings and work orders on site[1].

NC AI also fine-tuned a welding model, VAETKI-Vision-3B-Weld, on a small domain dataset while keeping the general vision skills intact. Compared with the base model, weld classification accuracy on RIAWELC rose from 41.1 to 80.9, defect detection and state inference on Khvatskii went from 39.3 to 69.8, and weld position and shape recognition on RefWeld jumped from 5.3 to 72.5[1][2].

NC AI says it built the pre-training and two-stage fine-tuning recipe on its own infrastructure, so that Korea's key industries and defence sector don't depend on outside technology. The company already works with Hanwha Ocean on autonomous welding robots, with POSCO DX on humanoid robot trials in industrial sites, and with Hyundai Rotem on world models[1].

Next, it plans to add reinforcement learning on top of the model and deploy it as the core vision engine and VLA base for robots and edge devices across its physical AI projects[1][2].

All numbers come from NC AI's own announcement, and the company hasn't published a paper or evaluation protocol alongside it. It also hasn't said whether the weights will be released, though earlier VAETKI language and vision models went up on Hugging Face under the NC AI consortium account[3].

Robots can't always call a big model in the cloud, and the move to vision-language-action models puts a premium on small vision models that understand space well enough to drive real-time control. If NC AI's results hold, a 3B model that beats 8B rivals on embodied reasoning, and can be cheaply tuned for jobs like weld inspection, is exactly the kind of component industrial robot makers need.