Gemma 4 Hits the Real World: Apache License, Edge Devices, and 140 Languages

News

Gemma 4 Launches Under Apache 2.0 With Four Model Sizes and 256K Context

Google dropped Gemma 4 with a real open-source license this time, Apache 2.0, which removes the legal friction that made earlier Gemma models awkward to ship commercially. The 31B dense model ranks #3 on Arena AI's open model leaderboard and fits on a single H100, while the 26B MoE activates only 3.8B parameters at inference for fast local throughput. This is the most deployable version of Gemma yet for builders who need frontier reasoning without cloud dependency.

Gemma 4 Becomes the Base Model for Gemini Nano 4 on Android, Claims 4x Speed and 60% Less Battery

Google is positioning Gemma 4 E2B and E4B as the production path for on-device Android AI, with native tool-calling support baked in for Agent Mode in Android Studio. The 4x speed improvement and 60% battery reduction over the previous generation are the numbers that will matter most to app developers shipping features to real users. If you're building Android apps with local inference, the AICore Developer Preview is worth picking up now to get ahead of the Gemini Nano 4 production launch.

Tools

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

vLLM getting day-zero support across every major hardware backend simultaneously is a signal that the open model deployment stack is maturing fast. If you're running heterogeneous GPU infrastructure, this means Gemma 4 works out of the box without waiting for vendor-specific patches. Practically useful for anyone building multi-tenant inference stacks today.