Deploy any robot AI model to any edge GPU. One command.

FastCrest builds Reflex, the open-source deployment layer for Vision-Language-Action (VLA) models — pi0, pi0.5, SmolVLA, NVIDIA's GR00T, OpenVLA. We collapse a month of CUDA / ONNX / TensorRT glue into one command (pip install reflex-vla && reflex go --model smolvla-base) with verified bit-for-bit numerical parity to the PyTorch reference.

Robotics teams training VLAs hit a wall at deployment. Most spend a month writing CUDA, ONNX, and TensorRT glue to get a single checkpoint running on a Jetson. Even after that, the export usually breaks numerically without anyone noticing until the robot misbehaves on real hardware. That silent action-drift is the most expensive class of bug in the field. Reflex eliminates it.

• Verified cos=+1.000000 ONNX parity across SmolVLA, pi0, pi0.5, GR00T N1.6, OpenVLA at machine precision. Five architectures, one tool.
• 9.79× speedup on decomposed pi0.5 via VLM-prefix + expert-denoise split with KV cache reuse.
• 5.55× speedup from TensorRT by default in v0.7+; 4.44× more from CUDA graphs on A100, p99 latency down 30%, jitter 4.1× tighter.
• First public reproduction of SnapFlow 1-NFE distillation — pi0.5 student scores 29/30 on LIBERO, beats the 10-step teacher's 28/30, fits an 8 GB Jetson Orin Nano.
• Continuous self-distilling serve ($99/mo Pro tier) hot-reloads a customer-specific 1-step student every few hours from real /act traffic.

• Site: https://fastcrest.com
• Discord: https://discord.gg/wrPUdcxdPu
• PyPI: https://pypi.org/project/reflex-vla/
• Email: team@fastcrest.com

BSL 1.1 — same model HashiCorp, MongoDB, Sentry, and Cockroach use. Free for personal, internal, and customer-facing commercial use; restricts only competing hosted/embedded offerings. Auto-converts to Apache 2.0 in four years.

Building toward the deployment layer of the robotics era. Software wedge → hardware bundles → silicon → datacenter. Read the full vision on Discord.