> getting started
Gemma 4 is Google DeepMind's open-weights model family, released April 2, 2026 under Apache 2.0. Five sizes run the same weights everywhere from a phone to a multi-GPU server, all multimodal (text plus image, with audio on the smaller models) and multilingual across 140+ languages. The first decision is which size fits your hardware; everything else is just which runtime you point at it.
// getting started
5 commands// faq
Is Gemma 4 free for commercial use?
Yes. Gemma 4 is released under Apache 2.0, which permits commercial use, modification, and redistribution with only attribution and license-notice requirements. This is a shift from Gemma 1 through 3, which used Google's custom Gemma Terms of Use. You run the weights yourself, so there is no per-token API cost.
What sizes does Gemma 4 come in?
Five: E2B (2.3B effective) and E4B (4.5B effective) for edge and mobile, a 12B unified multimodal model, a 26B A4B Mixture-of-Experts that activates 3.8B parameters per token, and a 31B dense model. E2B and E4B have a 128K context window; the 12B, 26B, and 31B reach 256K.
How much VRAM do I need to run Gemma 4?
Weights-only, roughly: E2B ~3 GB at 4-bit, E4B ~5 GB, 12B ~7 GB, 26B A4B ~15 GB, and 31B ~17 GB. In bf16 those roughly triple (31B is ~58 GB). Add several GB for the KV cache at long context. The E2B 4-bit build runs on CPU or a small GPU with no accelerator required.
What is the difference between the E models and the 12B/26B/31B?
The E2B and E4B are edge models that use Per-Layer Embeddings to keep active compute far below the loaded parameter count, so they fit tiny memory budgets and add audio input. The 12B, 26B, and 31B are larger, higher-quality models with a 256K context; the 26B is a Mixture-of-Experts tuned for throughput, and the 31B dense is the most capable overall.
Do I need a Google account or API key to use Gemma 4?
No. Gemma 4 is open weights: pull it from Ollama, Hugging Face, or Kaggle and run it locally with no key. Some Hugging Face repos still ask you to acknowledge the license in the browser before the first download, but there is no metered API involved when you self-host.