Show HN: InstinctFlash – Run 5B world-action models in real time on Jetson Thor

InstinctFlash, a high-performance serving framework for robotics models, has been released under the AGPL-3.0 license. It provides speedups of 1.2x to 7.9x on Jetson Thor when combined with a distilled few-step diffusion scheduler, and achieves a 90.5% success rate on 50 Robotwin2.0 tasks. The framework currently supports 8 VLA/world-action model families, including pi0.5 and NVIDIA Cosmos Policy, and can be used with fine-tuned checkpoints and exposed through a Python runtime or OpenPI-compatible WebSocket server.