Google expands LiteRT to microcontrollers and browsers now

What happened: Google published new Gemma documentation and model releases in December 2025 and emphasizes LiteRT (the lightweight on-device runtime) for constrained platforms including browsers, embedded Linux devices, and microcontrollers (sources: 1, 2). The FunctionGemma family includes a 270M-parameter model fine-tuned for mobile actions and on-device function calling, explicitly targeted at low-latency agent use cases (sources: 2, 3).

Technical details: LiteRT (formerly branded TensorFlow Lite) already provides a microcontroller-focused runtime whose core can fit in about 16 KB on an Arm Cortex‑M3. It is available as a C++17 library and as an Arduino package for TinyML projects (source: 4). Google’s Gemma pages document model conversion and optimization paths to prepare models for on-device execution within the LiteRT ecosystem (source: 1).

Implications: For teams building privacy-sensitive or offline agent features, combining compact Gemma models with a small-footprint runtime lowers latency and reduces cloud round trips and data egress. The 16 KB microcontroller runtime claim is documented in the LiteRT microcontrollers guide and should be validated on target hardware and toolchains before production deployment (source: 4).

Why It Matters

• Deploy compact agents: FunctionGemma’s 270M variant targets on-device function calling and can lower latency and data egress versus cloud-only agents.
• TinyML enabled: LiteRT’s microcontroller runtime (core ≈16 KB on Cortex‑M3) enables basic ML on devices with kilobytes of RAM for sensor preprocessing and local triggers.
• Lower integration work: Official Arduino and C++17 libraries plus model conversion docs shorten the path from model to device for embedded teams.
• Validate on target hardware: The microcontroller size claim appears only in Google’s LiteRT docs and must be benchmarked per board and build configuration before production.