MODELS

Google: Gemma 4 26B A4B

by Google DeepMind

modellead-sourceopenrouter-modelsgemmagoogle-deepmindmixture-of-expertsinstruction-tunedopen-modelopenroutersource:ai.google.dev

Visit Website Documentation

Overview

Gemma 4 26B A4B IT is an instruction-tuned Mixture-of-Experts model from Google DeepMind with 25.2B total parameters and 3.8B active per token during inference.

Details

Gemma 4 26B A4B IT is described by OpenRouter as an instruction-tuned Mixture-of-Experts model from Google DeepMind. Google’s Gemma 4 launch post says Gemma 4 was released in E2B, E4B, 26B Mixture of Experts, and 31B Dense sizes, and that the 26B MoE activates 3.8B parameters during inference. The official Google DeepMind Gemma 4 page lists Gemma 4 26B A4B IT Thinking in benchmarks and links model weights on Hugging Face, Ollama, Kaggle, LM Studio, and Docker.

When to Use

Use when you want to evaluate Google DeepMind’s Gemma 4 26B Mixture-of-Experts open model variant for instruction-tuned tasks. Use when lower active-parameter inference is important: the supplied Google source states the 26B MoE activates 3.8B parameters during inference. Use when you need availability through model-weight distribution channels listed by the official Gemma 4 page including Hugging Face Ollama Kaggle LM Studio and Docker.

Getting Started

Read the official Gemma 4 page at Google DeepMind for model context
benchmarks
and linked weight locations.
Review Google’s Gemma 4 launch post for the released model sizes and the 26B MoE active-parameter description.
If using OpenRouter
inspect the OpenRouter model page for provider-specific access details for google/gemma-4-26b-a4b-it.
Run a small evaluation against your own prompts before relying on the model in production.

Key Features

•Instruction-tuned Gemma 4 26B A4B IT model.
•Mixture-of-Experts architecture according to the OpenRouter and Google launch excerpts.
•25.2B total parameters with 3.8B active per token during inference
•per supplied source text.
•Official Gemma 4 page lists model weights on Hugging Face
•Ollama
•Kaggle
•LM Studio
•and Docker.
•Official Google DeepMind page includes Gemma 4 26B A4B IT Thinking in benchmarks.

Capabilities

•instruction-tuned text generation
•Mixture-of-Experts inference
•model-weight access through listed distribution platforms

Last updated Jun 1, 2026