Gemma 4 Edge Agentic Runtime: LiteRT-LM, AICore, and On-Device Agent Skills Explained

Tools

Gemma 4 Agent Skills Run Fully On-Device via LiteRT-LM and AI Edge Gallery

LiteRT-LM runs Gemma 4 E2B in under 1.5GB RAM with 2-bit/4-bit weights, and the new Agent Skills feature in AI Edge Gallery lets you build multi-step agentic workflows entirely offline on iOS and Android. On a Qualcomm Dragonwing IQ8, you get 3,700 prefill tokens per second, which makes this a serious local inference option, not a demo toy.

LiteRT-LM Python Package and CLI Now Available for Edge and IoT AI Pipelines

Google shipped a new Python package and CLI tool so you can run Gemma 4 on Linux, macOS, and Raspberry Pi without writing any inference code from scratch. Tool calling is supported out of the box in the CLI, which means you can wire up agentic pipelines on low-power hardware today.

Gemma 4 in Android Studio Agent Mode: Hardware Requirements and Setup Guide

The 26B MoE is the recommended model for most Android developers and needs 24GB total RAM with 17GB storage, while the E2B model works on 8GB machines. If you are evaluating whether your current workstation can run local agentic coding workflows, this table gives you the exact numbers to check against.

Gemma 4 on Cloud Run with NVIDIA Blackwell RTX PRO 6000: Serverless Inference That Scales to Zero

Cloud Run now supports Gemma 4 inference on NVIDIA RTX PRO 6000 GPUs with 96GB vGPU memory, scale-to-zero billing, and no reservations required in us-central1 and europe-west4. For teams that want serverless GPU inference without managing Kubernetes, this is the fastest path to production today.

Analysis

Gemma 4 Runs on Raspberry Pi 5 and Qualcomm Dragonwing IQ8 with Real Benchmark Numbers

Google published concrete performance numbers: Raspberry Pi 5 CPU hits 133 prefill and 7.6 decode tokens per second, while the Dragonwing IQ8 NPU jumps to 3,700 prefill tokens per second. This is the kind of grounded data that lets you actually plan an IoT or edge deployment rather than guess.

News

Gemma 4 AICore Developer Preview: Forward-Compatibility Path for Android Production Apps

Android developers can now prototype agentic flows using Gemma 4 via the AICore Developer Preview, which is designed to be forward-compatible with Gemini Nano 4 when it ships to production devices. Building on AICore now means your app architecture does not need to change when the system model upgrades.