Google Gemma 4 — Apache 2.0, MoE, Audio, and 256K Context in a 26B Model

Google's Gemma 4 (April 2, 2026) rewrites the open-source model tier: four sizes from E2B to 31B, a 26B MoE variant active at only 4B params per token, Per-Layer Embeddings for edge efficiency, audio input on the smaller models, 256K context on the larger ones — and a true Apache 2.0 license. GPQA Diamond: 84.3%. LiveCodeBench: 80.0%. AIME 2026: 89.2%. Community downloaded it 2 million times in weeks.