Gemma 4 on Android, Agentic Edge Skills, and the Apache 2.0 Advantage for Builders

Tools

Gemma 4 Sets the Baseline for Local Agentic AI on Android

Google is positioning Gemma 4 as the foundation for the next Gemini Nano generation, promising 4x faster inference and 60% less battery drain on Android hardware. For mobile app developers, the AICore Developer Preview is the practical on-ramp to prototype agentic features now before Gemini Nano 4 ships on flagship devices later this year.

Agent Skills in Google AI Edge Gallery: Multi-Step Autonomous Workflows Fully On-Device

Google's AI Edge Gallery now runs multi-step agentic workflows entirely on-device using Gemma 4, including Wikipedia queries, data visualizations, and cross-model integrations with no server required. LiteRT-LM hits under 1.5GB memory for E2B and processes 4,000 tokens across two skills in under 3 seconds, which are real numbers that matter for production mobile apps.

vLLM Adds Day-Zero Gemma 4 Support Across NVIDIA, AMD, Intel, and Google TPUs

vLLM shipping day-zero support across four hardware backends simultaneously is a sign the open model deployment ecosystem has matured significantly. Builders who already run vLLM in production can drop in Gemma 4 today without waiting for backend patches or compatibility fixes.

News

Gemma 4 Official Launch: Four Model Sizes, Apache 2.0, and 400M Downloads Worth of Momentum

Google shipped Gemma 4 in four sizes ranging from phone-class E2B to a 31B dense model that ranks third among open models globally, all under a fully permissive Apache 2.0 license. The license change is the real story here as it removes the last friction point for commercial deployment and fine-tuning without data leaving your infrastructure.

Analysis

Honest Gemma 4 Breakdown: Hardware Requirements, Cost Math, and Where It Actually Loses to Hosted Models

This review does the cost math clearly: the 26B MoE variant runs on a 12-16GB VRAM GPU and costs nothing per token after hardware, making it financially competitive with $20/month hosted subscriptions at moderate-to-high usage volumes. The honest limitations section is worth reading too since complex multi-constraint prompts and frontier-level reasoning tasks still favor hosted models.