South Korean artificial intelligence company NC AI has developed a lightweight visual language model, VAETKI Vision 3B, with approximately 3 billion parameters. It achieved an average score of 63.8 in three benchmark tests related to robotics and spatial understanding, outperforming six globally available models selected by the company. The relevant technology targets physical artificial intelligence in industries such as manufacturing, shipbuilding, and defense, aiming to enable robots and industrial equipment to autonomously understand their surrounding environment.
The primary function of the visual language model is to identify objects and spatial relationships within images and videos, and to comprehend their meanings. For robots to perform tasks based on human instructions, they must first assess the positions, movements, and environmental conditions of surrounding objects. NC AI's approach is to miniaturize the model, as robots and production equipment do not have the ample power and computational capacity of servers, making it difficult to directly deploy large models.
VAETKI Vision 3B Excels in Benchmark Tests
The three benchmarks used for testing include BLINK, which examines multi-image spatial relationships; ERIQ, which assesses the understanding of robotic actions and surrounding contexts; and ERQA, which infers object positions and spatial relationships. In a conventional mode that provides answers directly without complex reasoning, VAETKI Vision 3B achieved an average score of 63.8, surpassing six publicly available models with parameter counts ranging from 2 billion to 8 billion, including Alibaba's Tongyi Qianwen and Mistral. In a step-by-step reasoning mode, it scored an average of 63.2, 14.8 points higher than the 48.4 scored by the 8 billion parameter MiniStral 3 8B Reasoning.
It should be noted that these figures are based on specific evaluation conditions published by the company and do not represent the model's performance across all visual language tasks. The model also achieved the highest scores among comparison models in the ChartQA and TextVQA tasks, which can be used to analyze dashboards, blueprints, and work instructions in industrial settings. Additionally, it can process videos of up to 4 seconds and 8 frames in sync, determining object movements and positional changes.

