What is Gemma 4?

Google DeepMind's open-weight model family released May 2026 under the Apache 2.0 license, built from Gemini 3 research. Gemma 4 includes four variants: E2B (2.3B active, ~155 tok/s, runs on iPhone), E4B (4B effective, ~125 tok/s, multimodal+audio), 26B-A4B (MoE with 128 experts, 3.8B active, Arena AI #6), and 31B Dense (Arena AI #3, strongest open model under 100B). All variants support 256K context, text+image input, and native function calling. The E2B/E4B models use Per-Layer Embeddings (PLE) and also accept audio input. According to the LLMCheck index, Gemma 4 26B-A4B scores 67/100 on the leaderboard, the highest of any model.