InfoQ 中文T2·Tools
72 pts
Google has launched the LiteRT Tensor API, a lightweight C++ interface that lets developers author edge model graphs operator by operator, running the same code across CPU, GPU and NPU without per-hardware rewrites. It supports direct weight import from Hugging Face safetensors, zero-copy buffer interop, and AI coding agents that generate and optimize graphs; in Google's demo, identical SAM 2 code ran about 1.5x faster than Apple MLX on a MacBook Pro.