According to the official announcement, Guangdong Zhidong Weilai, in collaboration with Professor Cai Panpan's team from Shanghai Chuangzhi Academy, has unveiled the MINT-4B multimodal VLA (Vision-Language-Action) large model. The model ranks among the top three in global benchmark evaluations of mainstream general-purpose robotics models, outperforming industry-standard models such as OpenVLA, GR00T, π, and UniVLA in technical metrics.

The core innovation of the MINT series lies in "replicating task intent rather than mechanically replicating trajectories." It pioneers the SDAT multiscale frequency-domain tokenization technique, enabling hierarchical decoding through cross-scale autoregressive inference. This technology is already integrated into the Xiao Zhi S2 humanoid robot, deployed across scenarios including scientific research, education, and commercial exhibitions, with the model having been commercially rolled out in multiple regions nationwide.