Google Gemma 4

Gemma 4 landed on 2 April 2026 in five sizes. E2B and E4B are the tiny ones, quoted at 2.3B and 4.5B effective parameters and aimed at phones and embedded hardware. Then a 12B unified model, a 26B A4B sparse mixture-of-experts that activates roughly 3.8B parameters per token and a 31B dense at the top. Everything takes text and images. E2B, E4B and 12B also take audio, covering speech recognition and speech translation. Context runs 128K on the two small models and 256K from the 12B up, with over 140 languages supported.

The licence change is what separates this generation from the last three. Gemma 1 through 3n shipped under the Gemma Terms of Use with a Prohibited Use Policy that Google could revise on its own, an obligation to pass those restrictions down to anyone using your derivative and a widely disputed reading under which models trained on Gemma-generated synthetic data fell inside Gemma's terms. Gemma 4 drops all of it for plain Apache 2.0. Earlier Gemma models and the specialised variants stay on the old terms, so check which generation a checkpoint comes from before you build on it.

For a working engineer the appeal is coverage. One family spans a 2B that runs on a handset and a 31B that runs on a workstation, with the same tokeniser family and prompt conventions, plus purpose-built siblings for embeddings, safety classification, translation and function calling. Deployment is well trodden: llama.cpp, Ollama, vLLM, MLX and LM Studio all took it on release, and thirteen hosted providers serve it if you want an endpoint instead. The gap is transparency. Google publishes weights and a model card, not the corpus or the training pipeline.