Google's lightweight runtime for on-device ML, recently rebranded as LiteRT. By 2026, it is a primary option for deploying compact models to mobile, embedded, a
LiteRT (formerly TensorFlow Lite) is Google's on-device inference runtime: quantized models running on phones, microcontrollers, and embedded NPUs with hardware delegates, the production standard for mobile/edge ML across billions of devices.
Pricing: Free and open-source.
Scope: it's positioned as the multi-framework on-device runtime (not TF-only), reflecting how mobile inference outgrew its TensorFlow origins while keeping the deployment toolchain.
Small quantized models, yes: on-device assistants and classifiers ship with it. Larger on-device LLMs often use specialized stacks; vision/audio remain LiteRT's heartland.