mavis-ai/Gemma4-26B-MoE

VISIONConcurrency Cost:2Model Size:26BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 11, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

mavis-ai/Gemma4-26B-MoE provides the exact, unmodified base weights of Google's Gemma 4 26B Mixture of Experts (MoE) model. This 26 billion parameter model, with a 32768 token context length, is designed for general-purpose reasoning and serves as a powerful foundation for local AI applications. It is particularly optimized for integration into multi-agentic AI systems and complex local inference workflows.

Loading preview...

Model Overview

This repository hosts the unmodified base weights of Google's Gemma 4 26B MoE (Mixture of Experts) model. It is a 26 billion parameter model with a 32768 token context length, distributed under the Apache License 2.0. The weights are identical to the official release, with no fine-tuning, quantization, or structural modifications applied.

Key Characteristics

  • Architecture: Gemma 4 26B MoE, a Mixture of Experts model from Google.
  • Parameter Count: 26 billion parameters.
  • Context Length: Supports a substantial 32768 token context window.
  • Unmodified Weights: Provides the exact base weights as released by Google, ensuring consistency with the original model's capabilities.
  • License: Distributed under the Apache License 2.0, allowing for broad use and redistribution.

Primary Use Cases

  • Local AI Workflows: Suitable for MLX or other local AI inference setups.
  • Reasoning Engine: Designed to serve as a robust local reasoning engine for complex AI applications.
  • Multi-Agentic Systems: Specifically highlighted for its role in projects like R.E.V.I.S., a local Cognitive OS for Multi-Agentic AI, enabling tasks like recursive web research, dynamic RAG generation, and multi-step logic.