Real robot loops served on FlashRT, landing next week. Every tile opens a full scene.
import flash_rt
model = flash_rt.load_model(
checkpoint="/path/to/checkpoint",
framework="torch", # or "jax"
autotune=3, # 0=off, 3=default, 5=thorough
)
actions = model.predict(
images=[base_img, wrist_img],
prompt="pick up the red block",
)Runtime code, model paths, kernels, demos, and upcoming serving material will live here first.
Open FlashRTFrom camera to motor in one runtime — a tight perception-to-action loop built for embodied systems.
Perceive, reason and act, closed into a single low-latency path.
Graph-replay execution holds every tick to a tight, predictable budget.
Planner, world model and actor share one runtime, handing off by buffer.
FlashRT runs the whole physical-AI stack — not just one family.
π0 · π0.5 · GR00T N1.x · OpenVLA
Cosmos · Wan video diffusion · Motus
Qwen3 · Qwen3.6 · long-context agents
Streaming ASR ‖ VLA · text-to-speech
Flow / diffusion action heads, CFG rollout
Image · video · proprioception fusion
A white-box C++ runtime of hand-tuned CUDA kernels — FP8 / NVFP4 quant and fused megakernels that beat vendor libraries on the shapes that matter.
One whole-graph runtime across data-center and edge silicon — what it covers today, and what's next.
RTX 5090 · Thor · Orin
Jetson-class, low-power control loops
CDNA / RDNA, MI-class accelerators
Expanding the white-box kernel layer
FlashRT manages state, not tokens.