VFlash: Parallel Vision Model for Physical AI and Robotics

Jim Fan introduces VFlash, a System One Vision Model designed to parse the visual world in parallel for Physical AI applications. The model outputs entire geometric shapes at once rather than decoding single tokens sequentially, enabling up to a 12x faster action decoding and a 5.1x overall robot inference speedup. According to the announcement, the approach achieves up to 61% lower 3D trajectory error by avoiding the cascading errors of autoregressive decoding. Open article →