Gemma 4 26B A4B IT

Gemma 4 26B A4B IT is an instruction-tuned model from Google DeepMind designed as part of the open Gemma 4 family, with weights published on Hugging Face under an Apache 2.0 license. It uses a Mixture-of-Experts architecture with roughly 25.2B total parameters but only about 3.8B activated per token, an efficiency-oriented design that aims to deliver denser-model quality at a lower per-token compute cost. This sparse-activation approach makes it well suited for developers who want a capable open-weight model that can run with lighter inference budgets than comparable dense checkpoints.

The model is built for multimodal assistants and agent-style applications: it accepts text, image, and video input (including short video clips) while producing text output, and it supports native function calling, configurable thinking or reasoning mode, and structured outputs. A long context window makes it practical for document-heavy or multi-turn reasoning workflows, and the open-weight release lets teams fine-tune, self-host, or integrate the model into existing pipelines. Practically, it fits use cases such as grounded visual question answering, tool-using chatbots, and reasoning-heavy text tasks where an open, Apache-licensed Google model is preferred over proprietary alternatives.