You train the model. Ixian makes it fast on the board it ships on, and keeps it fast as batteries drain, firmware updates, and hardware ages.
Runs in your environment or ours. Your weights never leave your infrastructure.
Same model, same board, same task. The clock runs until the robot moves. Codex and Claude Code are what an in-house engineer would reach for, so that's the bar.
Time to first action: from camera frame ready to the first actuator command leaving the model. Not model load, not warmup. The loop your robot actually lives in.
Jetson AGX Thor devkit, MAXN power mode, active cooling, JetPack 7.0. Same physical board for every column.
500 timed runs after a 50-run warmup; we report the median. Batch 1, FP16→INT8 mixed precision only where the model's own eval suite confirms the accuracy delta shown.
Codex CLI 0.21 and Claude Code 1.9 were each given the same brief (make this model as fast as possible on this board) with unlimited attempts. Best verified result kept.
Jetson AGX Thor · MAXN · FP16→INT8 mixed · batch 1 · JetPack 7.0 · n=500 · 2026-08-27 · Codex CLI 0.21 · Claude Code 1.9 · latency = ms per action chunk · PRELIMINARY
Full run logs are available to design partners. The optimization pipeline itself is ours.
Every retrain and every JetPack or TensorRT release quietly undoes hand-tuned work. Ixian re-optimizes and re-benchmarks on every trigger, before the fleet feels it.
Thermal paste degrades. Storage wears. OS updates change scheduling. An on-robot probe compares live p99 and energy against the day-one baseline, then tells you which unit to fix and why.
the fleet console · simulation
Scroll. One model becomes five intelligence levels, each pruned and compressed to a different degree. RoboRouter dispatches every task to the level the moment calls for.
Inference isn't a step. It's a loop. We close it.
You ship the model once. Ixian keeps it optimized, routed, and watched, then re-optimizes the moment anything changes.
yeonseok@zetic.ai