Frequently Asked Questions

Gemma 4 is released under the Apache 2.0 open-source license. This provides complete developer flexibility — you can use it commercially, modify it, and deploy it anywhere without restrictive barriers.

The E2B and E4B models run on mobile devices (Android phones, Raspberry Pi, NVIDIA Jetson Orin Nano). The 26B MoE and 31B Dense models run on consumer gaming GPUs (quantized) or a single 80GB H100/A100 GPU (fp16).

Gemma 4 is an open-weights model you can download and run on your own hardware, built from the same research as Gemini 3. Gemini is Google's proprietary model accessed via API. Gemma 4 is ideal for privacy, offline use, and fine-tuning.

Gemma 4 31B ranks #3 on the Arena AI open-source leaderboard and outperforms models 20x its size. While closed models like GPT-4o may still lead on raw capability, Gemma 4 offers unmatched performance-per-parameter among models you can run locally.

The 26B MoE (Mixture of Experts) activates only 3.8B parameters per inference, making it much faster with lower latency. The 31B Dense activates all parameters, providing the highest raw quality and being the best base for fine-tuning.

Yes. All Gemma 4 models natively process images and video. The E2B and E4B edge models additionally support native audio input for speech recognition and understanding.

Gemma 4 is natively trained on over 140 languages, scoring 85.2% on MMMLU (multilingual question answering). The training data spans diverse linguistic content for strong multilingual understanding.

Model weights are available on Hugging Face (google/gemma-4 collection), Kaggle, Ollama (ollama pull gemma4), LM Studio, and Docker Hub (ai/gemma4). You can try it instantly in Google AI Studio (26B and 31B).