The domestic video deep reasoning model VLX-VR has been released
On September 1, Hangzhou Lianhui Technology Co., Ltd. officially released the video reasoning model VLX-VR. This model ranked first on the video reasoning benchmark MINERVA jointly developed by Google DeepMind and Columbia University, with an accuracy rate of 78.8% in video问答 tasks, surpassing international flagship models such as Google Gemini 3.5 and OpenAI GPT-4.1. VLX-VR achieves a logic consistency of 96.2%, and its accuracy increases not decreases with time for video lengths ranging from 5 to 15 minutes, reaching a maximum of 80.92%. This release is a continuation of OmAI Lianhui’s technical achievements; previously, in February 2025, the company opened-sourced the visual reasoning model VLM-R1 on GitHub, introducing the强化 learning reasoning framework of DeepSeek-R1 into the visual domain for the first time. As a fundamental ability for physical AI to perceive the world, progress in video reasoning capabilities coincides with the accelerated standardization process of China’s robotics industry.