prithivMLmods/Qwen3.5-4B-Unredacted-MAX

VISIONConcurrency Cost:1Model Size:4.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Mar 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

prithivMLmods/Qwen3.5-4B-Unredacted-MAX is a 4.5 billion parameter language model based on the Qwen3.5-4B architecture, optimized for improved loading stability and enhanced compatibility with modern Transformers pipelines. This version maintains the reasoning and instruction-following behavior of its base model, focusing on efficient inference and lightweight deployment. It is designed for research-oriented experimentation and local AI prototyping, particularly for understanding transformer behavior and instruction dynamics.

Loading preview...

Qwen3.5-4B-Unredacted-MAX Overview

This model is an optimized release of the huihui-ai/Huihui-Qwen3.5-4B-abliterated base, featuring a 4.5 billion parameter architecture. It has been re-sharded and packaged for improved loading stability and enhanced compatibility with current Hugging Face Transformers versions. The model preserves the core reasoning and instruction-following capabilities of its base, making it suitable for efficient deployment and research.

Key Capabilities & Features

  • Optimized Packaging: Improved repository structure for smoother downloads and inference initialization.
  • Stable Transformers Compatibility: Designed for modern Hugging Face Transformers workflows.
  • 4B Parameter Architecture: Lightweight and resource-efficient, ideal for local inference and experimentation.
  • Improved Instruction Handling: Maintains consistent behavior across structured prompts and multi-step instructions.
  • High Abliteration Rate: Self-reported 90.5% non-refusal rate, indicating a focus on direct response generation.

Good For

  • Research into transformer behavior and instruction-following dynamics.
  • Lightweight local AI deployment and prototyping.
  • Red-teaming and robustness testing of language models.
  • Efficient inference on hardware with limited resources.