In real-robot tests conducted on September 15, 2026, the InstinctFlash framework demonstrated up to a 33.78× speedup on NVIDIA's Jetson Thor compared to baseline measurements, while maintaining task performance. The framework manages model declarations, optimization planning, and runtime execution through an inspectable pipeline.
InstinctFlash: High-Performance Serving Framework for Robotics Models
InstinctFlash supports various model families, including LingBot-VA, pi0.5, GR00T, and NVIDIA's Cosmos3. The runtime can apply optimizations such as quantization and specific scheduling changes to enhance execution efficiency. It also offers a Python API for local integration and a WebSocket-based protocol compatible with the openpi ecosystem.
The framework includes evaluation tools to compare optimized models against original versions. These tools report latency, action agreement, and simulator task success to assist in performance assessment.
Sources
- Show HN: InstinctFlash – Run 5B world-action models in real time on Jetson Thor （Hacker News Frontpage, 2026-09-22）