Gemma 4 26B A4B by Google DeepMind — 256K Context

25.2B total parameter (3.8B active) Mixture-of-Experts model with 256K token context. 8 active experts out of 128 total + 1 shared expert. Runs inference speed comparable to a 4B dense model while matching larger dense model quality. Supports text, image, and video input with text output, configurable thinking, native function calling, and multilingual support.