Google AI Edge has released LiteRT-LM 0.15, adding a new Apple Foundation Models adapter to its on-device large language model runtime. The release notes say the adapter connects Apple's backend with the LiteRT-LM runtime for text and multimodal execution, including vision and audio paths using Gemma 4 models.

The update matters because LiteRT-LM is Google's cross-platform orchestration layer for running LLMs with LiteRT. The project README describes it as a production-ready runtime layer for high-performance execution across device classes, and says it powers on-device generative AI experiences in Chrome, Chromebook Plus, Pixel, and Android.

Version 0.15 also adds CLI configuration support. Developers can now define global defaults and per-model settings in a centralized config.json, covering choices such as hardware backends, thinking budgets, and context window sizes. The release notes define parameter precedence as runtime arguments first, then model-specific config, global defaults, model metadata, and engine fallbacks.

On the web side, the release adds Gemma 4 support for LiteRT-LM.js, including 12B, 26B A4B, and 31B web model files, plus a new use_autosized_ringbuffers option for long-context memory efficiency.

This is an incremental runtime release rather than a new product launch, but it is a concrete step in Google's edge AI stack. It widens the places where the same LiteRT-LM runtime can sit: native Apple-backed execution, browser-based Gemma 4 runs, and configured local CLI workflows.