Native Model Authoring for Edge with LiteRT Tensor API- Google Developers Blog

Introduces the LiteRT Tensor API, a new C++ authoring layer for handcrafting on-device AI models. It provides full control over graph structure, direct weight ingestion from checkpoints (e.g., .safetensors), tensor-centric buffer interoperability for zero-copy data transfer on accelerators (CPU, GPU, NPU), and serves as a foundation for AI coding agents. Demonstrates its use by authoring the Segment Anything Model 2 (SAM 2) in C++ via the API, achieving real-time performance on WebGPU and showing cross-platform performance on Metal (macOS, iOS) and OpenCL (Android). The API supports exporting graphs as .tflite files for LiteRT execution or direct CPU execution via XNNPACK.