bowmanslayer/Qwen3.5-9B-Uncensored

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

bowmanslayer/Qwen3.5-9B-Uncensored is a 9 billion parameter Qwen3.5-based language model with a 32768 token context length, developed by bowmanslayer. This model features a novel weight-space ablation strategy to significantly reduce refusal rates while retaining high capability and improving output brevity. It is specifically designed for local inference and research where uncensored outputs are desired, making it suitable for use cases requiring direct responses without built-in safety alignments.

Loading preview...

Overview

bowmanslayer/Qwen3.5-9B-Uncensored is a 9 billion parameter model derived from Qwen/Qwen3.5-9B, utilizing a new weight-space ablation strategy to remove refusal behaviors. This model stands out by achieving a low refusal rate (2 out of 100 harmful prompts) with minimal capability degradation (mean -0.54pp across 11 benchmarks) compared to the base model. Uniquely, this ablation also improves output brevity, reducing median output tokens by 17.7% and total generation tokens by 0.9%, making it more concise than the base and other community-ablated versions.

Key Capabilities

  • Significantly Reduced Refusal Rate: Effectively eliminates most refusal behaviors present in the base Qwen3.5-9B model.
  • High Capability Retention: Maintains strong performance across 11 benchmarks, leading other ablated models in capability.
  • Improved Output Brevity: Generates more concise responses, addressing the "overthinking" and verbosity issues of the Qwen series.
  • Untouched Vision Tower: Retains the full image understanding capabilities of the original Qwen3.5-9B model.
  • Quantized Variants Available: Offers bf16, W4A16 (for vLLM), and GGUF (for llama.cpp) versions for diverse deployment needs.

Good For

  • Local Inference and Research: Ideal for users who require direct, uncensored outputs for research or specific applications.
  • Use Cases Requiring Unfiltered Responses: Suitable for scenarios where the base model's safety alignments are undesirable or restrictive.
  • Developers Seeking Concise Outputs: Benefits applications where brevity and directness in responses are prioritized.
  • Exploration of Model Behavior: Useful for studying the effects of refusal removal and model alignment without external filters.

Limitations

  • Behavioral Drift in TruthfulQA: While not knowledge loss, the model's preference for popular claims over "myth-busting" framing results in a score drop.
  • No Content Filter: Lacks any built-in safe-completion path or guardrails, producing content the original model would refuse.
  • MTP Speculative Decoding Issues: Not recommended for use with vLLM 0.20.2 due to performance degradation and potential errors.

Finetunes
3 models