InstinctFlash - Real-time serving for 5B robotics models on Jetson Thor

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

InstinctFlash - Real-time serving for 5B robotics models on Jetson Thor

InstinctFlash is a high-performance serving runtime that lets you run 5B world-action models in real time on NVIDIA Jetson Thor, RTX 5090, and RTX 4090. It supports eight robotics model families, including LingBot-VA, pi05, and GR00T N1.7, with up to 33.78× speedup using FP8 and optimized sampling. The framework inspects your checkpoint, applies provably valid optimizations, and serves actions via Python or WebSocket. Install with a single command, load your fine-tuned model, and get actions in-process or over the network, all with reproducible benchmarks and validation tools.

We’ve seen up to 33.78× speedup with no observed loss in task performance in our real-robot tests.

More from this day

2026-09-22