Google released version 0.15.0 of LiteRT-LM, its open-source framework for running language models directly on phones, laptops and small devices. The update adds an adapter for Apple's own on-device machine learning backend, a single configuration file for the command-line tool instead of many flags, and browser support for Gemma 4 models. It also fixes crashes that occurred when several model engines were started at the same time.
What changed
Earlier versions had no Apple Foundation backend adapter and required model parameters to be passed as command-line flags.
What it unlocks
Running Gemma 4 models on Apple devices through Apple's own backend, and setting model parameters once in a config.json file instead of long command lines.
- v0.15.0
- Gemma 4 web support up to 31B
- Multi-Token Prediction up to 3x faster inference
What you need to act on it
- building from source or installing the CLI
- a supported platform such as Android, iOS, web, desktop or Raspberry Pi
- github.com2026-08-05