mewse/gemma-4-26B-A4B-it-qat-q4_0-unquantized-heretic-ara

Model Overview

This model, mewse/gemma-4-26B-A4B-it-qat-q4_0-unquantized-heretic-ara, is a 26 billion parameter instruction-tuned version of Google DeepMind's Gemma 4 A4B model. It has been decensored using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) to significantly reduce content refusal rates. While the original model had 100/100 refusals in evaluation, this Heretic-modified version achieves a refusal rate of 5/100 (or 10/100 at IQ4_XS quantization), demonstrating a substantial reduction in restrictive outputs.

Key Capabilities

• Decensored Output: Engineered to provide less restrictive content generation compared to its base model, with a significantly lower refusal rate.
• Multimodal: Processes both text and image inputs, with support for variable aspect ratios and resolutions.
• Reasoning: Designed with configurable thinking modes for step-by-step problem-solving.
• Extended Context: Features a 256K token context window, enabling complex, long-context tasks.
• Efficient Architecture: As a Mixture-of-Experts (MoE) model, it has 25.2B total parameters but only 3.8B active parameters, allowing for faster inference comparable to a 4B model.
• Coding & Agentic Workflows: Shows improved performance in coding benchmarks and includes native function-calling support.

Good For

• Applications requiring less restrictive content: Ideal for use cases where the base model's high refusal rate is a limitation.
• Multimodal tasks: Excels in scenarios involving interleaved text and image inputs, such as image understanding, document parsing, and UI analysis.
• Reasoning and complex problem-solving: Benefits from built-in reasoning modes and a large context window.
• Code generation and agentic systems: Supports advanced coding tasks and function calling for autonomous agents.
• Efficient deployment: Its MoE architecture allows for faster inference on consumer GPUs and workstations despite its large total parameter count.