Gemma 4 26B A4B IT is the mixture-of-experts member of the Gemma 4 family that Google DeepMind introduced as part of its 2026 open-weights lineup, released under the Apache 2.0 license and shipped in both pre-trained and instruction-tuned forms. The family spans five sizes, and the 26B A4B variant uses the MoE design that Gemma 4 was built to showcase, sitting alongside dense options so deployers can trade scale against efficiency. As an open release with a public launch blog and a technical report on arXiv, the model is positioned for community fine-tuning and on-device-to-server deployment rather than as a closed API product.

The instruction-tuned 26B A4B variant is designed for reasoning-heavy workloads, with configurable thinking modes and a multimodal profile that handles text and image input while emitting text output, supported by a context window of up to the cataloged API limit tokens and multilingual coverage across more than 140 languages. The family description frames it as well suited to text generation, coding, and reasoning tasks, and the MoE architecture lets it serve those use cases with relatively lean active compute compared with the dense members. In practice it is a fit for developers who want an open, long-context multimodal model that can be self-hosted or pulled through standard model hubs without proprietary licensing restrictions.