nDimensional/Qwen3.5-9B-Uncensored-Safetensors

VISIONConcurrency Cost:1Model Size:9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Mar 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

nDimensional/Qwen3.5-9B-Uncensored-Safetensors is a 9 billion parameter Qwen3.5 dense hybrid attention model, converted to Safetensors format from HauhauCS's uncensored GGUF version. This model retains the Qwen3.5-9B architecture, including its vision capabilities, but incorporates modifications made by HauhauCS for uncensoring. It is designed for use in frameworks like vLLM and transformers that require Safetensors, offering an uncensored Qwen3.5 variant with a 32K context length.

Loading preview...

Model Overview

This model, nDimensional/Qwen3.5-9B-Uncensored-Safetensors, is a 9 billion parameter variant of the Qwen3.5 architecture, specifically converted to the Safetensors format. It originates from HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive, which is an uncensored version of the base Qwen/Qwen3.5-9B model. The conversion process ensures compatibility with frameworks that do not natively support GGUF or split GGUF formats, such as vLLM and Hugging Face Transformers.

Key Characteristics

  • Architecture: Qwen3.5 dense hybrid attention, featuring 24 GDN and 8 standard attention layers.
  • Parameters: 9 billion, with a context length of 32,768 tokens.
  • Uncensored Nature: Incorporates modifications from HauhauCS aimed at reducing censorship.
  • Vision Capabilities: The vision encoder weights are directly copied from the official Qwen3.5-9B, ensuring full visual processing capabilities are retained.
  • Safetensors Format: Provided in BF16 Safetensors, making it readily usable in various deep learning frameworks.
  • Conversion Accuracy: Tensor analysis confirms that core components like the tokenizer, embeddings, and MLP layers are bit-identical to the official Qwen3.5-9B, with noted modifications in self_attn.o_proj.weight and linear_attn.norm.weight reflecting the uncensoring efforts.

Ideal Use Cases

  • Framework Compatibility: Developers requiring a Qwen3.5-9B model in Safetensors format for vLLM, Hugging Face Transformers, or similar environments.
  • Uncensored Applications: Scenarios where a less restrictive language model is preferred.
  • Vision-Language Tasks: Applications benefiting from the Qwen3.5's integrated vision encoder for multimodal understanding.
  • Research and Development: Exploring the impact of uncensoring modifications on model behavior and performance.

Finetunes
1 models