Gemma 4 for Private Local AI: Apache 2.0, Edge Hardware, and Agentic Coding in One Release

News

Gemma 4 Drops the Restrictive License: Full Apache 2.0 for Enterprise and Edge

The real story here isn't the model capabilities, it's the licensing shift. Apache 2.0 means you can redistribute, embed, and commercialize without the old Gemma Terms of Use friction, which is a genuine unlock for OEMs, healthcare deployments, and anyone building products on top of local AI.

Tools

Android Studio Gets Gemma 4 Local Model Support with Full Agent Mode

Gemma 4 now powers Agent Mode in Android Studio, running entirely on your machine via Ollama or LM Studio with no API key required. The 26B MoE variant handles multi-file refactoring, build error resolution, and feature generation, making this the most capable offline coding assistant Google has shipped.

Gemma 4 on vLLM with Day 0 Support Across TPU, AMD, and Intel Hardware

vLLM ships Gemma 4 with Day 0 support on Google TPUs, AMD GPUs, and Intel XPUs alongside Nvidia, which is a meaningful expansion of the hardware surface for production serving. The 128K to 256K context windows and native function calling make this a serious option for agentic serving workloads beyond just local dev.

LiteRT-LM Runs Gemma 4 E2B Under 1.5GB RAM on Mobile and Raspberry Pi

LiteRT-LM gets Gemma 4 running on Raspberry Pi 5 at 133 prefill tokens per second on CPU, and jumps to 3,700 on a Qualcomm Dragonwing NPU. Sub-1.5GB RAM usage on mobile with 2-bit and 4-bit weights plus tool calling support makes this a real option for IoT and robotics builders, not just a demo.

Analysis

Gemma 4 on NVIDIA Jetson and DGX Spark: Production Deployment Guide for Edge AI

NVIDIA covers the full deployment stack from Jetson Orin Nano running E2B and E4B at the edge to DGX Spark running the full 31B in BF16 with vLLM. The NVFP4 quantized checkpoint for Blackwell is the detail worth tracking for anyone optimizing cost per token in production.