google/gemma-4-26B-A4B-it-qat-q4_0-unquantized

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The google/gemma-4-26B-A4B-it-qat-q4_0-unquantized model is a multimodal Mixture-of-Experts (MoE) variant from the Gemma 4 family by Google DeepMind, optimized with Quantization-Aware Training (QAT). This 25.2 billion total parameter model, with 3.8 billion active parameters, handles text and image input, generating text output, and features a 256K token context window. It excels in reasoning, coding, and agentic workflows, offering efficient inference comparable to a 4B parameter model.

Loading preview...

Gemma 4 26B A4B MoE: Multimodal, Efficient, and Reasoning-Capable

This model is a Gemma 4 family member developed by Google DeepMind, featuring a Mixture-of-Experts (MoE) architecture. It is optimized with Quantization-Aware Training (QAT), allowing for significantly reduced memory requirements while maintaining quality. The model processes text and image inputs, generating text outputs, and supports a 256K token context window.

Key Capabilities

  • Multimodal: Processes text and image inputs, with variable aspect ratio and resolution support for images. It can analyze video by processing sequences of frames.
  • Efficient Architecture: As a 25.2 billion total parameter MoE model, it activates only 3.8 billion parameters during inference, providing performance comparable to a 4B parameter model.
  • Reasoning: Designed with configurable thinking modes for step-by-step reasoning.
  • Enhanced Coding & Agentic Capabilities: Achieves notable improvements in coding benchmarks and includes native function-calling support for autonomous agents.
  • Multilingual: Supports over 140 languages in pre-training and 35+ languages out-of-the-box.
  • Native System Prompt Support: Allows for more structured and controllable conversations.

Good For

  • Fast Inference: Ideal for scenarios requiring efficient processing due to its MoE architecture's low active parameter count.
  • Complex Reasoning Tasks: Benefits from its built-in reasoning mode and strong performance on benchmarks like MMLU Pro and AIME 2026.
  • Multimodal Applications: Suitable for tasks involving interleaved text and image inputs, such as object detection, document parsing, and UI understanding.
  • Coding and Agentic Workflows: Excels in code generation, completion, correction, and structured tool use.

Finetunes
2 models