ApxInf v0.1

ApxInf v0.1

The first release of ApxInf, a Rust-based inference engine focused on embodied AI, delivers optimized PI0.5 inference on NVIDIA Jetson Thor, Orin, and RTX GPUs.

Supported models

• PI0.5: BF16, FP8, and INT8 inference, with hardware-specific optimizations.
• WallOSS: Native VLA inference with Torch-free preprocessing.
• Llama and Qwen3-VL: Built-in text and vision-language inference.

Performance highlights — PI0.5

Hardware	Precision	Action generation	Latency	Frequency
Jetson AGX Thor	FP8	10-step	41.16 ms	24.3 Hz
Jetson AGX Orin	BF16	10-step	165.67 ms	6.0 Hz
RTX 4090	INT8	10-step	25.99 ms	38.5 Hz
Jetson AGX Thor	FP8	One-step	26.32 ms	38.0 Hz
Jetson AGX Orin	BF16	One-step	119.05 ms	8.4 Hz

Measured as steady-state CUDA Graph replay P50, batch size 1, two 224×224 views, and action horizon 10. One-step results use action generation pruning.

• LIBERO-10 success rates: 94.0% on Thor FP8, 92.8% on Thor BF16, and 92.0% on Orin BF16, over 500 episodes each.
• Includes Python policy APIs, OpenPI-compatible WebSocket serving, FP8 calibration, kernel autotuning, and an agent-assisted model-porting workflow.