LiteRT Tensor API

Developers can use it to author and optimize portable on-device model pipelines, including preprocessing and postprocessing, with shared accelerator buffers.

Availability

Tensor API guide, reference, and buildable SAM 2 sample are available

Pricing model

Not established

Deployment

On-device; LiteRT on CPU, GPU, or NPU, or directly on CPU through XNNPACK

Privacy

On-device execution

Licence

Not established

• Google announced the release of the LiteRT Tensor API. ^[1]

  To solve the above challenges, we are excited to announce the release of the novel LiteRT Tensor API , a lightweight authoring layer in the LiteRT stack, designed to improve handcrafting models.

• Developers can build computation graphs op by op in C++. ^[1]
• The API can load weights directly from checkpoints such as Hugging Face .safetensors files. ^[1]

  Direct weight ingestion : load weights straight from model checkpoints, such as Hugging Face .safetensors files, without conversion scripts needed in between.

• The same graph can run through LiteRT on CPU, GPU, or NPU, or directly on CPU through XNNPACK. ^[1]

  The same graph can run through LiteRT on CPU, GPU (powered by open-source ML Drift ) and NPU, or without model file or runtime directly on CPU through XNNPACK, so you don't rewrite it for each new hardware target.

• Preprocessing, model partitions, and postprocessing can share accelerator buffers without copying intermediate data. ^[1]

  Tensor-centric buffer interop : pre-processing, model partitions, and post-processing pass tensors to each other through shared accelerator buffers, so intermediate data stays on the accelerator with zero copies.

• Google provides a Tensor API guide, reference, and buildable end-to-end SAM 2 example. ^[1]

  Point it to the Tensor API README and the SAM 2 sample ; together they cover the API and a buildable end-to-end example.

• OpenGCUA for macOS leads 12 AI developments on 19 October 2026 Daily brief 2026-10-19