Google Ships Gemma 4, Edge-Optimized Open Models With 31B Dense Variant

Google Ships Gemma 4 Open Models

Google released Gemma 4 on April 2, 2026. The family is released under the Apache 2.0 license and is available for commercial use.

The Four Variants

According to Google, Gemma 4 comes in four sizes: E2B, E4B, 26B A4B, and 31B Dense. Google describes the E2B and E4B variants as edge-optimized, activating an effective 2 billion and 4 billion parameters during inference to preserve RAM and battery life on constrained devices. The 31B Dense model is the flagship for higher-end workloads.

Google describes the variants as multimodal, supporting text, image, and video input, with support for over 140 languages and context windows up to 256K tokens. Specific per-variant features should be checked on Google's official model pages.

Performance Claims

Google reports that the 31B Dense model ranks highly among open models on public leaderboards. Google also highlights strong gains on competitive coding benchmarks compared to Gemma 3. Specific benchmark numbers, including Codeforces ELO figures that have circulated in press coverage, should be sourced directly from Google's announcement or model card before citation.

Edge Deployment

Google promotes the smaller Gemma 4 variants for on-device use. Google's Android Developers Blog highlights support for local agentic workflows through AICore and mentions developer scenarios on devices like NVIDIA Jetson Orin Nano. Not every variant is intended for every class of device, so developers should match the variant size to their hardware.

Licensing

The models are released under the Apache 2.0 license. There are no monthly active user limits in the license, no special acceptable use policy beyond standard license terms, and no permission process required for commercial use.

Why It Matters

Gemma 4 continues a steady pattern of stronger open models from Google. For mobile apps, private AI assistants, and offline workflows, the edge variants give developers a serious open option. For heavier workloads, the 31B Dense model is an open alternative worth benchmarking against closed options.