Google Launches Gemma 4: Its Most Intelligent Open-Source AI Model Yet

Google has launched Gemma 4, its most capable open-source AI model family to date, built on the same technology that underpins Gemini 3. Available under the Apache 2.0 license — one of the most permissive in open-source software — Gemma 4 is purpose-built for advanced reasoning and agentic workflows, and runs on everything from enterprise cloud clusters to a Raspberry Pi.

Four Model Variants, One Big Claim

Gemma 4 ships in four sizes: a 31-billion parameter dense model, a 26-billion parameter mixture-of-experts (MoE) variant that only activates 3.8B parameters at inference time, and two lightweight edge models — E4B and E2B. The larger models support a 256K token context window; the edge variants offer 128K.

On Arena AI's global text leaderboard for open models, the 31B dense ranks 3rd and the 26B MoE ranks 6th — both outperforming models 20 times their size. Google's pitch: "byte for byte, the most capable open models" available today.

Built for Agents, Runs Anywhere

Gemma 4 includes native function calling and structured JSON output, making it drop-in ready for agentic pipelines. The smaller E2B and E4B models also support audio input, enabling voice-driven agents on-device with near-zero latency. Google confirms full compatibility with Android phones, Jetson Orin Nano, and Arm-based hardware — real-time autonomous decision-making without a cloud dependency.

On the server side, the 26B MoE model's architecture activates only a fraction of its parameters per token, delivering speed close to a 4B model at 26B-quality output — ideal for cost-sensitive production deployments.

Why Apache 2.0 Is a Big Deal

Previous Gemma releases used a custom Google license that left commercial use in a legal grey area. Switching to Apache 2.0 removes that ambiguity entirely. Developers can now build and sell products on Gemma 4, modify it freely, distribute derivatives, and do so with patent protection — no strings attached, no risk of Google retroactively tightening terms. This directly challenges Meta's Llama series, which uses its own custom license, and positions Gemma 4 as the most commercially open major model family available.

Where to Get It

Model weights are available now on Hugging Face, Kaggle, and via Ollama for one-command local deployment. Cloud access is available through Google AI Studio (free tier), Vertex AI, and Google Cloud. NVIDIA RTX GPU users can accelerate inference via CUDA and Tensor Core optimizations announced alongside the launch.

The Bottom Line

Gemma 4 makes a strong case for being the go-to open-source model for teams that need commercial freedom, edge deployment, and competitive reasoning — all without the size overhead of frontier models. For developers who have been waiting for a genuinely permissive, production-grade open model, the wait is over.