Understand the Gemma 4 model family

Why watch

Google DeepMind's Maarten Grootendorst introduces the Gemma 4 family of open source multimodal language models, covering five model sizes across four architectures including the efficient E2B and E4B with per-layer embeddings, the encoder-free 12B, the 26B mixture of experts, and the most capable 31B dense model. This overview is designed for developers and AI practitioners looking to understand the Gemma 4 ecosystem and choose the right model for their use case.