Google released LiteRT Tensor API on Nov. 3, 2026. The lightweight C++ interface lets developers build, customize, and optimize machine-learning models for on-device execution, allowing them to define computation graphs operation by operation for CPU, GPU, and NPU targets without backend rewrites. The API supports direct weight ingestion from checkpoints, including Hugging Face .safetensors files, and zero-copy buffer interoperability that keeps accelerator buffers intact across pipeline stages. Coding agents can generate, modify, validate, and benchmark the C++ graphs. Google demonstrated the workflow with a coding agent that built a multi-object segmentation and tracking application using the SAM 2 model, executing identical C++ code on a MacBook Pro running Metal, an iPhone running Metal, and an Android phone running OpenCL; on the MacBook Pro, the implementation ran about 1.5 times faster than Apple MLX in fp16 mode. Arm director of AI Laurence Moroney wrote on LinkedIn that developers should start with model conversion and reserve the C++ API for cases requiring graph-level control or accelerator-bound pre- and post-processing, highlighting that performance on WebGPU at 512x512 resolution jumped from under 10 frames per second to over 30. Google is working on direct ML Drift backend support and additional NPU-friendly operations, and plans automatic full-graph quantization and expanded dynamic-shape support.