General Instinct launches InstinctFlash to accelerate robotics AI inference on edge hardware

General Instinct, cofounded by Guanming, has released InstinctFlash, an open-source (AGPL-3.0) high-performance inference framework designed to speed up robotics AI models on edge devices. Running on NVIDIA's Jetson Thor platform, the framework delivers runtime speedups of 1.2x to 7.9x, and up to 33.78x for the LingBot-VA model when combining runtime optimizations with a distilled few-step diffusion scheduler. Benchmarked across 1,153 episodes per configuration on 50 RoboTwin 2.0 tasks, the accelerated LingBot-VA achieved a 90.5% success rate, slightly below the 92.1% baseline but at a fraction of the computational cost. InstinctFlash supports eight vision-language-action and world-action model families, including pi0.5 and NVIDIA Cosmos Policy, across consumer and professional GPUs as well as Jetson Thor. Companies including Samsung and Siemens have already used the framework, and it is now publicly available on GitHub.